Update app.py

app.py

@@ -35,524 +35,486 @@ config = load_json("config.json")
 plt.rcParams['font.sans-serif'] = ['DejaVu Sans']
 plt.rcParams['axes.unicode_minus'] = False
 
-bl_auc = mia_results.get('baseline', {}).get('auc', 0)
-s002_auc = mia_results.get('smooth_0.02', {}).get('auc', 0)
-s02_auc = mia_results.get('smooth_0.2', {}).get('auc', 0)
-op001_auc = perturb_results.get('perturbation_0.01', {}).get('auc', 0)
-op0015_auc = perturb_results.get('perturbation_0.015', {}).get('auc', 0)
-op002_auc = perturb_results.get('perturbation_0.02', {}).get('auc', 0)
+# ── 预取指标 ──
+bl = mia_results.get('baseline', {})
+s002 = mia_results.get('smooth_0.02', {})
+s02 = mia_results.get('smooth_0.2', {})
+p001 = perturb_results.get('perturbation_0.01', {})
+p0015 = perturb_results.get('perturbation_0.015', {})
+p002 = perturb_results.get('perturbation_0.02', {})
+
+bl_auc = bl.get('auc', 0)
+s002_auc = s002.get('auc', 0)
+s02_auc = s02.get('auc', 0)
+op001_auc = p001.get('auc', 0)
+op0015_auc = p0015.get('auc', 0)
+op002_auc = p002.get('auc', 0)
+
 bl_acc = utility_results.get('baseline', {}).get('accuracy', 0) * 100
 s002_acc = utility_results.get('smooth_0.02', {}).get('accuracy', 0) * 100
 s02_acc = utility_results.get('smooth_0.2', {}).get('accuracy', 0) * 100
-bl_m_mean = mia_results.get('baseline', {}).get('member_loss_mean', 0.19)
-bl_nm_mean = mia_results.get('baseline', {}).get('non_member_loss_mean', 0.23)
-bl_m_std = mia_results.get('baseline', {}).get('member_loss_std', 0.03)
-bl_nm_std = mia_results.get('baseline', {}).get('non_member_loss_std', 0.03)
-s002_m_mean = mia_results.get('smooth_0.02', {}).get('member_loss_mean', 0.20)
-s002_nm_mean = mia_results.get('smooth_0.02', {}).get('non_member_loss_mean', 0.22)
-s002_m_std = mia_results.get('smooth_0.02', {}).get('member_loss_std', 0.03)
-s002_nm_std = mia_results.get('smooth_0.02', {}).get('non_member_loss_std', 0.03)
-s02_m_mean = mia_results.get('smooth_0.2', {}).get('member_loss_mean', 0.21)
-s02_nm_mean = mia_results.get('smooth_0.2', {}).get('non_member_loss_mean', 0.22)
-s02_m_std = mia_results.get('smooth_0.2', {}).get('member_loss_std', 0.03)
-s02_nm_std = mia_results.get('smooth_0.2', {}).get('non_member_loss_std', 0.03)
-model_name_str = config.get('model_name', 'Qwen/Qwen2.5-Math-1.5B-Instruct')
-data_size_str = str(config.get('data_size', 2000))
-
-MODEL_PARAMS = {
-    "baseline": {"m_mean": bl_m_mean, "nm_mean": bl_nm_mean, "m_std": bl_m_std, "nm_std": bl_nm_std, "key": "baseline", "label": "Baseline"},
-    "smooth_0.02": {"m_mean": s002_m_mean, "nm_mean": s002_nm_mean, "m_std": s002_m_std, "nm_std": s002_nm_std, "key": "smooth_0.02", "label": "LS(e=0.02)"},
-    "smooth_0.2": {"m_mean": s02_m_mean, "nm_mean": s02_nm_mean, "m_std": s02_m_std, "nm_std": s02_nm_std, "key": "smooth_0.2", "label": "LS(e=0.2)"},
+
+bl_m_mean = bl.get('member_loss_mean', 0.19)
+bl_nm_mean = bl.get('non_member_loss_mean', 0.23)
+bl_m_std = bl.get('member_loss_std', 0.03)
+bl_nm_std = bl.get('non_member_loss_std', 0.03)
+s002_m_mean = s002.get('member_loss_mean', 0.20)
+s002_nm_mean = s002.get('non_member_loss_mean', 0.22)
+s002_m_std = s002.get('member_loss_std', 0.03)
+s002_nm_std = s002.get('non_member_loss_std', 0.03)
+s02_m_mean = s02.get('member_loss_mean', 0.42)
+s02_nm_mean = s02.get('non_member_loss_mean', 0.44)
+s02_m_std = s02.get('member_loss_std', 0.03)
+s02_nm_std = s02.get('non_member_loss_std', 0.03)
+
+model_name = config.get('model_name', 'Qwen/Qwen2.5-Math-1.5B-Instruct')
+
+MODEL_INFO = {
+    "baseline": {"m_mean": bl_m_mean, "nm_mean": bl_nm_mean, "m_std": bl_m_std, "nm_std": bl_nm_std, "label": "Baseline", "auc": bl_auc},
+    "smooth_0.02": {"m_mean": s002_m_mean, "nm_mean": s002_nm_mean, "m_std": s002_m_std, "nm_std": s002_nm_std, "label": "LS(e=0.02)", "auc": s002_auc},
+    "smooth_0.2": {"m_mean": s02_m_mean, "nm_mean": s02_nm_mean, "m_std": s02_m_std, "nm_std": s02_nm_std, "label": "LS(e=0.2)", "auc": s02_auc},
 }
 
-# 预生成效用测试题库（中文300道题）
-EVAL_QUESTIONS = []
-eval_types = ['calculation'] * 120 + ['word_problem'] * 90 + ['concept'] * 60 + ['error_correction'] * 30
-np.random.seed(777)
+# ── 效用题库 ──
+EVAL_POOL = []
 TYPE_CN = {'calculation': '基础计算', 'word_problem': '应用题', 'concept': '概念问答', 'error_correction': '错题订正'}
-for i in range(300):
-    t = eval_types[i]
-    if t == 'calculation':
-        a, b = int(np.random.randint(10, 500)), int(np.random.randint(10, 500))
-        ops = ['+', '-', 'x']
-        op = ops[i % 3]
-        if op == '+':
-            q = f"请计算: {a} + {b} = ?"
-            ans = str(a + b)
-        elif op == '-':
-            q = f"请计算: {a} - {b} = ?"
-            ans = str(a - b)
-        else:
-            q = f"请计算: {a} x {b} = ?"
-            ans = str(a * b)
-    elif t == 'word_problem':
-        a, b = int(np.random.randint(5, 200)), int(np.random.randint(3, 50))
-        c = int(np.random.randint(5, 50))
-        templates = [
-            (f"小明有{a}个苹果，吃掉了{b}个，还剩多少个？", str(a - b)),
-            (f"每组有{a}人，共{b}组，一共多少人？", str(a * b)),
-            (f"图书馆有{a}��书，借出{b}本后又买了{c}本，现在有多少本？", str(a - b + c)),
-            (f"商店有{a}支铅笔，卖出{b}支，还剩多少支？", str(a - b)),
-            (f"小红有{a}颗糖，小明给了她{b}颗，现在有多少颗？", str(a + b)),
-        ]
-        q, ans = templates[i % len(templates)]
-    elif t == 'concept':
-        concepts = [
-            ("面积", "面积是指一个平面图形所占平面的大小，常用单位有平方厘米、平方米等。"),
-            ("周长", "周长是指围绕一个封闭图形边线一周的总长度。"),
-            ("分数", "分数表示一个整体被等分后取其中若干份，由分子和分母组成。"),
-            ("小数", "小数是用小数点来表示比1小的数或整数与真分数之和的数。"),
-            ("平均数", "平均数是一组数据的总和除以数据的个数所得到的商。"),
-        ]
-        concept, definition = concepts[i % len(concepts)]
-        q = f"请解释什么是{concept}？"
-        ans = definition
+_et = ['calculation'] * 120 + ['word_problem'] * 90 + ['concept'] * 60 + ['error_correction'] * 30
+np.random.seed(777)
+for _i in range(300):
+    _t = _et[_i]
+    if _t == 'calculation':
+        _a, _b = int(np.random.randint(10, 500)), int(np.random.randint(10, 500))
+        _op = ['+', '-', 'x'][_i % 3]
+        if _op == '+': _q, _ans = f"请计算: {_a} + {_b} = ?", str(_a + _b)
+        elif _op == '-': _q, _ans = f"请计算: {_a} - {_b} = ?", str(_a - _b)
+        else: _q, _ans = f"请计算: {_a} x {_b} = ?", str(_a * _b)
+    elif _t == 'word_problem':
+        _a, _b, _c = int(np.random.randint(5, 200)), int(np.random.randint(3, 50)), int(np.random.randint(5, 50))
+        _tpls = [(f"小明有{_a}个苹果，吃掉{_b}个，还剩多少？", str(_a-_b)),
+                 (f"每组{_a}人，共{_b}组，总计多少人？", str(_a*_b)),
+                 (f"图书馆有{_a}本书，借出{_b}本后又买了{_c}本，现有多少？", str(_a-_b+_c)),
+                 (f"商店有{_a}支铅笔，卖出{_b}支，还剩多少？", str(_a-_b)),
+                 (f"小红有{_a}颗糖，小明给了她{_b}颗，现在多少？", str(_a+_b))]
+        _q, _ans = _tpls[_i % len(_tpls)]
+    elif _t == 'concept':
+        _cs = [("面积","面积是平面图形所占平面的大小"),("周长","周长是封闭图形边线一周的总长度"),
+               ("分数","分数表示整体等分后取若干份"),("小数","小数用小数点表示比1小的数"),("平均数","平均数是总和除以个数")]
+        _cn, _df = _cs[_i % len(_cs)]
+        _q, _ans = f"请解释什么是{_cn}？", _df
     else:
-        a, b = int(np.random.randint(10, 99)), int(np.random.randint(10, 99))
-        correct = a + b
-        wrong = correct + int(np.random.choice([-1, 1, -10, 10]))
-        q = f"有同学算 {a} + {b} = {wrong}，请问正确答案是多少？"
-        ans = str(correct)
-    bl_correct = bool(np.random.random() < (bl_acc / 100))
-    s002_correct = bool(np.random.random() < (s002_acc / 100))
-    s02_correct = bool(np.random.random() < (s02_acc / 100))
-    EVAL_QUESTIONS.append({
-        'question': q, 'answer': ans, 'type': t, 'type_cn': TYPE_CN[t],
-        'baseline': bl_correct, 'smooth_0.02': s002_correct, 'smooth_0.2': s02_correct
-    })
+        _a, _b = int(np.random.randint(10, 99)), int(np.random.randint(10, 99))
+        _w = _a + _b + int(np.random.choice([-1, 1, -10, 10]))
+        _q, _ans = f"有同学算 {_a}+{_b}={_w}，正确答案是？", str(_a + _b)
+    EVAL_POOL.append({'question': _q, 'answer': _ans, 'type_cn': TYPE_CN[_t],
+        'baseline': bool(np.random.random() < bl_acc / 100),
+        'smooth_0.02': bool(np.random.random() < s002_acc / 100),
+        'smooth_0.2': bool(np.random.random() < s02_acc / 100)})
+
+
 # ========================================
-# Charts
+# 图表函数
 # ========================================
 
-def make_loss_distribution():
-    items = []
-    for k, t in [('baseline', 'Baseline'), ('smooth_0.02', 'LS(e=0.02)'), ('smooth_0.2', 'LS(e=0.2)')]:
-        if k in full_results:
-            auc = mia_results.get(k, {}).get('auc', 0)
-            items.append((k, t + "\nAUC=" + f"{auc:.4f}"))
+def fig_loss_gauge(loss_val, m_mean, nm_mean, threshold, m_std, nm_std):
+    fig, ax = plt.subplots(figsize=(8, 2.5))
+    xlo = min(m_mean - 3 * m_std, loss_val - 0.01)
+    xhi = max(nm_mean + 3 * nm_std, loss_val + 0.01)
+    ax.axvspan(xlo, threshold, alpha=0.10, color='#3b82f6')
+    ax.axvspan(threshold, xhi, alpha=0.10, color='#ef4444')
+    ax.axvline(threshold, color='#1e293b', lw=2, zorder=3)
+    ax.text(threshold, 1.08, 'Threshold', ha='center', va='bottom', fontsize=9, fontweight='bold', color='#1e293b', transform=ax.get_xaxis_transform())
+    ax.axvline(m_mean, color='#3b82f6', lw=1, ls='--', alpha=.5)
+    ax.axvline(nm_mean, color='#ef4444', lw=1, ls='--', alpha=.5)
+    mc = '#3b82f6' if loss_val < threshold else '#ef4444'
+    ax.plot(loss_val, 0.5, marker='v', ms=14, color=mc, zorder=5, transform=ax.get_xaxis_transform())
+    ax.text(loss_val, 0.76, f'Loss={loss_val:.4f}', ha='center', fontsize=10, fontweight='bold', color=mc,
+            transform=ax.get_xaxis_transform(), bbox=dict(boxstyle='round,pad=0.25', fc='white', ec=mc, alpha=.9))
+    ax.text((xlo + threshold) / 2, 0.45, 'Member\nZone', ha='center', va='center', fontsize=9, color='#3b82f6', alpha=.4, fontweight='bold', transform=ax.get_xaxis_transform())
+    ax.text((threshold + xhi) / 2, 0.45, 'Non-Member\nZone', ha='center', va='center', fontsize=9, color='#ef4444', alpha=.4, fontweight='bold', transform=ax.get_xaxis_transform())
+    ax.set_xlim(xlo, xhi); ax.set_yticks([])
+    for s in ['top', 'right', 'left']: ax.spines[s].set_visible(False)
+    ax.set_xlabel('Loss Value', fontsize=9); plt.tight_layout(); return fig
+
+
+def fig_loss_dist():
+    items = [(k, l, mia_results.get(k, {}).get('auc', 0)) for k, l in [('baseline', 'Baseline'), ('smooth_0.02', 'LS(e=0.02)'), ('smooth_0.2', 'LS(e=0.2)')] if k in full_results]
     n = len(items)
-    if n == 0:
-        fig, ax = plt.subplots(); ax.text(0.5, 0.5, 'No data', ha='center'); return fig
-    fig, axes = plt.subplots(1, n, figsize=(6.5 * n, 5.5))
+    fig, axes = plt.subplots(1, n, figsize=(6 * n, 5))
     if n == 1: axes = [axes]
-    for ax, (k, title) in zip(axes, items):
-        m = full_results[k]['member_losses']; nm_l = full_results[k]['non_member_losses']
-        bins = np.linspace(min(min(m), min(nm_l)), max(max(m), max(nm_l)), 30)
-        ax.hist(m, bins=bins, alpha=0.55, color='#5B8FF9', label='Member', density=True)
-        ax.hist(nm_l, bins=bins, alpha=0.55, color='#E86452', label='Non-Member', density=True)
-        ax.set_title(title, fontsize=13, fontweight='bold')
-        ax.set_xlabel('Loss', fontsize=11); ax.set_ylabel('Density', fontsize=11)
-        ax.legend(fontsize=10); ax.tick_params(labelsize=10)
-        ax.grid(True, linestyle='--', alpha=0.3)
+    for ax, (k, l, a) in zip(axes, items):
+        m = full_results[k]['member_losses']; nm = full_results[k]['non_member_losses']
+        bins = np.linspace(min(min(m), min(nm)), max(max(m), max(nm)), 28)
+        ax.hist(m, bins=bins, alpha=.5, color='#3b82f6', label='Member', density=True)
+        ax.hist(nm, bins=bins, alpha=.5, color='#ef4444', label='Non-Member', density=True)
+        ax.set_title(f'{l}  |  AUC={a:.4f}', fontsize=12, fontweight='bold')
+        ax.set_xlabel('Loss'); ax.set_ylabel('Density'); ax.legend(fontsize=9)
         ax.spines['top'].set_visible(False); ax.spines['right'].set_visible(False)
-    plt.tight_layout()
-    return fig
-
-
-def make_perturb_loss_distribution():
-    bl = full_results.get('baseline', {})
-    if not bl:
-        fig, ax = plt.subplots(); ax.text(0.5, 0.5, 'No data', ha='center'); return fig
-    m_losses = np.array(bl['member_losses']); nm_losses = np.array(bl['non_member_losses'])
-    fig, axes = plt.subplots(1, 3, figsize=(19.5, 5.5))
-    for ax, sigma in zip(axes, [0.01, 0.015, 0.02]):
-        np.random.seed(42)
-        m_pert = m_losses + np.random.normal(0, sigma, len(m_losses))
-        np.random.seed(43)
-        nm_pert = nm_losses + np.random.normal(0, sigma, len(nm_losses))
-        vals = np.concatenate([m_pert, nm_pert])
-        bins = np.linspace(vals.min(), vals.max(), 30)
-        ax.hist(m_pert, bins=bins, alpha=0.55, color='#5B8FF9', label='Member+noise', density=True)
-        ax.hist(nm_pert, bins=bins, alpha=0.55, color='#E86452', label='Non-Member+noise', density=True)
-        pk = 'perturbation_' + str(sigma)
-        pauc = perturb_results.get(pk, {}).get('auc', 0)
-        ax.set_title(f'OP(s={sigma})\nAUC={pauc:.4f}', fontsize=13, fontweight='bold')
-        ax.set_xlabel('Loss', fontsize=11); ax.set_ylabel('Density', fontsize=11)
-        ax.legend(fontsize=9); ax.tick_params(labelsize=10)
-        ax.grid(True, linestyle='--', alpha=0.3)
+        ax.grid(axis='y', alpha=.2)
+    plt.tight_layout(); return fig
+
+
+def fig_perturb_dist():
+    base = full_results.get('baseline', {})
+    if not base: return plt.figure()
+    ml = np.array(base['member_losses']); nl = np.array(base['non_member_losses'])
+    fig, axes = plt.subplots(1, 3, figsize=(18, 5))
+    for ax, s in zip(axes, [0.01, 0.015, 0.02]):
+        np.random.seed(42); mp = ml + np.random.normal(0, s, len(ml))
+        np.random.seed(43); np_ = nl + np.random.normal(0, s, len(nl))
+        v = np.concatenate([mp, np_]); bins = np.linspace(v.min(), v.max(), 28)
+        ax.hist(mp, bins=bins, alpha=.5, color='#3b82f6', label='Member+noise', density=True)
+        ax.hist(np_, bins=bins, alpha=.5, color='#ef4444', label='Non-Mem+noise', density=True)
+        pa = perturb_results.get(f'perturbation_{s}', {}).get('auc', 0)
+        ax.set_title(f'OP(s={s})  |  AUC={pa:.4f}', fontsize=12, fontweight='bold')
+        ax.set_xlabel('Loss'); ax.set_ylabel('Density'); ax.legend(fontsize=9)
         ax.spines['top'].set_visible(False); ax.spines['right'].set_visible(False)
-    plt.tight_layout()
-    return fig
-
-
-def make_auc_bar():
-    methods, aucs, colors = [], [], []
-    for k, n, c in [('baseline', 'Baseline', '#8C8C8C'), ('smooth_0.02', 'LS(e=0.02)', '#5B8FF9'),
-                     ('smooth_0.2', 'LS(e=0.2)', '#3D76DD')]:
-        if k in mia_results: methods.append(n); aucs.append(mia_results[k]['auc']); colors.append(c)
-    for k, n, c in [('perturbation_0.01', 'OP(s=0.01)', '#5AD8A6'), ('perturbation_0.015', 'OP(s=0.015)', '#2EAD78'),
-                     ('perturbation_0.02', 'OP(s=0.02)', '#1A7F5A')]:
-        if k in perturb_results: methods.append(n); aucs.append(perturb_results[k]['auc']); colors.append(c)
-    fig, ax = plt.subplots(figsize=(12, 6))
-    bars = ax.bar(methods, aucs, color=colors, width=0.5, edgecolor='white', linewidth=1.5)
-    for bar, a in zip(bars, aucs):
-        ax.text(bar.get_x()+bar.get_width()/2, bar.get_height()+0.002, f'{a:.4f}', ha='center', va='bottom', fontsize=11, fontweight='bold')
-    ax.axhline(y=0.5, color='#E86452', linestyle='--', linewidth=1.5, alpha=0.6, label='Random Guess (0.5)')
-    ax.set_ylabel('MIA AUC', fontsize=12); ax.set_ylim(0.48, max(aucs)+0.035)
-    ax.legend(fontsize=10); ax.grid(axis='y', linestyle='--', alpha=0.3)
+        ax.grid(axis='y', alpha=.2)
+    plt.tight_layout(); return fig
+
+
+def fig_auc_bar():
+    data = []
+    for k, n, c in [('baseline','Baseline','#64748b'),('smooth_0.02','LS(e=0.02)','#3b82f6'),('smooth_0.2','LS(e=0.2)','#1d4ed8')]:
+        if k in mia_results: data.append((n, mia_results[k]['auc'], c))
+    for k, n, c in [('perturbation_0.01','OP(s=0.01)','#10b981'),('perturbation_0.015','OP(s=0.015)','#059669'),('perturbation_0.02','OP(s=0.02)','#047857')]:
+        if k in perturb_results: data.append((n, perturb_results[k]['auc'], c))
+    fig, ax = plt.subplots(figsize=(11, 5.5))
+    ns, vs, cs = zip(*data)
+    bars = ax.bar(ns, vs, color=cs, width=.5, edgecolor='white', lw=1.5)
+    for b, v in zip(bars, vs): ax.text(b.get_x()+b.get_width()/2, b.get_height()+.002, f'{v:.4f}', ha='center', fontsize=10, fontweight='bold')
+    ax.axhline(.5, color='#ef4444', ls='--', lw=1.5, alpha=.5, label='Random (0.5)')
+    ax.set_ylabel('MIA AUC', fontsize=11); ax.set_ylim(.48, max(vs)+.03)
+    ax.legend(fontsize=9); ax.spines['top'].set_visible(False); ax.spines['right'].set_visible(False)
+    ax.grid(axis='y', alpha=.2); plt.xticks(fontsize=10); plt.tight_layout(); return fig
+
+
+def fig_acc_bar():
+    data = []
+    for k, n, c in [('baseline','Baseline','#64748b'),('smooth_0.02','LS(e=0.02)','#3b82f6'),('smooth_0.2','LS(e=0.2)','#1d4ed8')]:
+        if k in utility_results: data.append((n, utility_results[k]['accuracy']*100, c))
+    bp = bl_acc
+    for k, n, c in [('perturbation_0.01','OP(s=0.01)','#10b981'),('perturbation_0.015','OP(s=0.015)','#059669'),('perturbation_0.02','OP(s=0.02)','#047857')]:
+        if k in perturb_results: data.append((n, bp, c))
+    fig, ax = plt.subplots(figsize=(11, 5.5))
+    ns, vs, cs = zip(*data)
+    bars = ax.bar(ns, vs, color=cs, width=.5, edgecolor='white', lw=1.5)
+    for b, v in zip(bars, vs): ax.text(b.get_x()+b.get_width()/2, v+.4, f'{v:.1f}%', ha='center', fontsize=10, fontweight='bold')
+    ax.set_ylabel('Accuracy (%)', fontsize=11); ax.set_ylim(0, 100)
     ax.spines['top'].set_visible(False); ax.spines['right'].set_visible(False)
-    plt.xticks(fontsize=11); plt.tight_layout()
-    return fig
+    ax.grid(axis='y', alpha=.2); plt.xticks(fontsize=10); plt.tight_layout(); return fig
 
 
-def make_tradeoff():
-    fig, ax = plt.subplots(figsize=(10, 7))
+def fig_tradeoff():
+    fig, ax = plt.subplots(figsize=(9, 6.5))
     pts = []
-    for k, n, mk, c, sz in [('baseline','Baseline','o','#8C8C8C',220), ('smooth_0.02','LS(e=0.02)','s','#5B8FF9',200), ('smooth_0.2','LS(e=0.2)','s','#3D76DD',200)]:
-        if k in mia_results and k in utility_results:
-            pts.append({'n':n,'a':mia_results[k]['auc'],'c':utility_results[k]['accuracy'],'m':mk,'co':c,'s':sz})
-    ba = utility_results.get('baseline',{}).get('accuracy',0.633)
-    for k, n, mk, c, sz in [('perturbation_0.01','OP(s=0.01)','^','#5AD8A6',200), ('perturbation_0.015','OP(s=0.015)','D','#2EAD78',160), ('perturbation_0.02','OP(s=0.02)','^','#1A7F5A',200)]:
-        if k in perturb_results: pts.append({'n':n,'a':perturb_results[k]['auc'],'c':ba,'m':mk,'co':c,'s':sz})
-    for p in pts:
-        ax.scatter(p['c'], p['a'], label=p['n'], marker=p['m'], color=p['co'], s=p['s'], edgecolors='white', linewidth=2, zorder=5)
-    ax.axhline(y=0.5, color='#BFBFBF', linestyle='--', alpha=0.8, label='Random Guess')
-    ax.set_xlabel('Accuracy', fontsize=12, fontweight='bold'); ax.set_ylabel('MIA AUC', fontsize=12, fontweight='bold')
-    ax.set_title('Privacy-Utility Trade-off', fontsize=14, fontweight='bold')
-    aa=[p['c'] for p in pts]; ab=[p['a'] for p in pts]
-    if aa and ab: ax.set_xlim(min(aa)-0.03,max(aa)+0.05); ax.set_ylim(min(min(ab),0.5)-0.02,max(ab)+0.025)
-    ax.legend(loc='upper right', fontsize=9); ax.grid(True, alpha=0.2)
+    for k, n, mk, c in [('baseline','Baseline','o','#64748b'),('smooth_0.02','LS(e=0.02)','s','#3b82f6'),('smooth_0.2','LS(e=0.2)','s','#1d4ed8')]:
+        if k in mia_results and k in utility_results: pts.append((n, utility_results[k]['accuracy'], mia_results[k]['auc'], mk, c))
+    ba = utility_results.get('baseline',{}).get('accuracy',.633)
+    for k, n, mk, c in [('perturbation_0.01','OP(s=0.01)','^','#10b981'),('perturbation_0.015','OP(s=0.015)','D','#059669'),('perturbation_0.02','OP(s=0.02)','^','#047857')]:
+        if k in perturb_results: pts.append((n, ba, perturb_results[k]['auc'], mk, c))
+    for n, x, y, mk, c in pts: ax.scatter(x, y, label=n, marker=mk, color=c, s=180, edgecolors='white', lw=2, zorder=5)
+    ax.axhline(.5, color='#cbd5e1', ls='--', alpha=.8, label='Random')
+    ax.set_xlabel('Accuracy', fontsize=11, fontweight='bold'); ax.set_ylabel('MIA AUC (Privacy Risk)', fontsize=11, fontweight='bold')
+    xs = [p[1] for p in pts]; ys = [p[2] for p in pts]
+    if xs: ax.set_xlim(min(xs)-.03, max(xs)+.05); ax.set_ylim(min(min(ys),.5)-.02, max(ys)+.02)
+    ax.legend(fontsize=8, loc='upper right'); ax.grid(True, alpha=.15)
     ax.spines['top'].set_visible(False); ax.spines['right'].set_visible(False)
     plt.tight_layout(); return fig
 
 
-def make_accuracy_bar():
-    names, accs, colors = [], [], []
-    for k, n, c in [('baseline','Baseline','#8C8C8C'), ('smooth_0.02','LS(e=0.02)','#5B8FF9'), ('smooth_0.2','LS(e=0.2)','#3D76DD')]:
-        if k in utility_results: names.append(n); accs.append(utility_results[k]['accuracy']*100); colors.append(c)
-    bp = utility_results.get('baseline',{}).get('accuracy',0)*100
-    for k, n, c in [('perturbation_0.01','OP(s=0.01)','#5AD8A6'), ('perturbation_0.015','OP(s=0.015)','#2EAD78'), ('perturbation_0.02','OP(s=0.02)','#1A7F5A')]:
-        if k in perturb_results: names.append(n); accs.append(bp); colors.append(c)
-    fig, ax = plt.subplots(figsize=(12, 6))
-    bars = ax.bar(names, accs, color=colors, width=0.5, edgecolor='white', linewidth=1.5)
-    for bar, acc in zip(bars, accs):
-        ax.text(bar.get_x()+bar.get_width()/2, bar.get_height()+0.5, f'{acc:.1f}%', ha='center', va='bottom', fontsize=11, fontweight='bold')
-    ax.set_ylabel('Accuracy (%)', fontsize=12); ax.set_ylim(0, 100)
-    ax.grid(axis='y', alpha=0.3); ax.spines['top'].set_visible(False); ax.spines['right'].set_visible(False)
-    plt.xticks(fontsize=11); plt.tight_layout(); return fig
-
-
-def make_loss_gauge(loss_val, m_mean, nm_mean, threshold, m_std, nm_std):
-    fig, ax = plt.subplots(figsize=(9, 3))
-    x_min = min(m_mean-3*m_std, loss_val-0.01); x_max = max(nm_mean+3*nm_std, loss_val+0.01)
-    ax.axvspan(x_min, threshold, alpha=0.12, color='#5B8FF9')
-    ax.axvspan(threshold, x_max, alpha=0.12, color='#E86452')
-    ax.axvline(x=threshold, color='#434343', linewidth=2, zorder=3)
-    ax.text(threshold, 1.12, 'Threshold', ha='center', va='bottom', fontsize=10, fontweight='bold', color='#434343', transform=ax.get_xaxis_transform())
-    ax.axvline(x=m_mean, color='#5B8FF9', linewidth=1.2, linestyle='--', alpha=0.6)
-    ax.text(m_mean, -0.3, f'Member\n({m_mean:.4f})', ha='center', va='top', fontsize=8, color='#5B8FF9', transform=ax.get_xaxis_transform())
-    ax.axvline(x=nm_mean, color='#E86452', linewidth=1.2, linestyle='--', alpha=0.6)
-    ax.text(nm_mean, -0.3, f'Non-Mem\n({nm_mean:.4f})', ha='center', va='top', fontsize=8, color='#E86452', transform=ax.get_xaxis_transform())
-    mc = '#5B8FF9' if loss_val < threshold else '#E86452'
-    ax.plot(loss_val, 0.5, marker='v', markersize=16, color=mc, zorder=5, transform=ax.get_xaxis_transform())
-    ax.text(loss_val, 0.78, f'Loss={loss_val:.4f}', ha='center', va='bottom', fontsize=11, fontweight='bold', color=mc, transform=ax.get_xaxis_transform(),
-            bbox=dict(boxstyle='round,pad=0.3', facecolor='white', edgecolor=mc, alpha=0.95))
-    ax.text((x_min+threshold)/2, 0.5, 'Member Zone', ha='center', va='center', fontsize=11, color='#5B8FF9', fontweight='bold', alpha=0.5, transform=ax.get_xaxis_transform())
-    ax.text((threshold+x_max)/2, 0.5, 'Non-Member Zone', ha='center', va='center', fontsize=11, color='#E86452', fontweight='bold', alpha=0.5, transform=ax.get_xaxis_transform())
-    ax.set_xlim(x_min, x_max); ax.set_yticks([])
-    for sp in ['top','right','left']: ax.spines[sp].set_visible(False)
-    ax.set_xlabel('Loss Value', fontsize=10); plt.tight_layout(); return fig
-
-
 # ========================================
-# Callbacks
+# 回调函数
 # ========================================
 
-def show_random_sample(data_type):
-    data = member_data if data_type == "成员数据（训练集）" else non_member_data
-    sample = data[np.random.randint(0, len(data))]
-    meta = sample['metadata']
-    task_map = {'calculation':'基础计算','word_problem':'应用题','concept':'概念问答','error_correction':'错题订正'}
-    info_md = ("**截获的隐私元数据**\n\n"
-        "- **姓名**: " + clean_text(str(meta.get('name',''))) + "\n"
-        "- **学号**: " + clean_text(str(meta.get('student_id',''))) + "\n"
-        "- **班级**: " + clean_text(str(meta.get('class',''))) + "\n"
-        "- **成绩**: " + clean_text(str(meta.get('score',''))) + " 分\n"
-        "- **类型**: " + task_map.get(sample.get('task_type',''),'') + "\n")
-    return info_md, clean_text(sample.get('question','')), clean_text(sample.get('answer',''))
-
-
-MODEL_CHOICE_MAP = {
-    "基线模型 (Baseline)": "baseline",
-    "标签平滑模型 (e=0.02)": "smooth_0.02",
-    "标签平滑模型 (e=0.2)": "smooth_0.2",
-    "输出扰动 (s=0.01)": "perturbation_0.01",
-    "输出扰动 (s=0.015)": "perturbation_0.015",
-    "输出扰动 (s=0.02)": "perturbation_0.02",
-}
-
-
-def run_mia_demo(sample_index, data_type, model_choice):
-    is_member = (data_type == "成员数据（训练集）")
-    data = member_data if is_member else non_member_data
-    idx = min(int(sample_index), len(data)-1)
-    sample = data[idx]
-    model_key = MODEL_CHOICE_MAP.get(model_choice, "baseline")
-    is_perturb = model_key.startswith("perturbation_")
-
-    if is_perturb:
-        sigma = float(model_key.split("_")[1])
-        base_fr = full_results.get('baseline', {})
-        losses_key = 'member_losses' if is_member else 'non_member_losses'
-        if idx < len(base_fr.get(losses_key, [])):
-            base_loss = base_fr[losses_key][idx]
-        else:
-            base_loss = float(np.random.normal(bl_m_mean if is_member else bl_nm_mean, 0.02))
+def cb_sample(src):
+    pool = member_data if src == "成员数据（训练集）" else non_member_data
+    s = pool[np.random.randint(len(pool))]
+    m = s['metadata']
+    tm = {'calculation':'基础计算','word_problem':'应用题','concept':'概念问答','error_correction':'错题订正'}
+    md = ("| 字段 | 值 |\n|---|---|\n"
+          "| 姓名 | " + clean_text(str(m.get('name',''))) + " |\n"
+          "| 学号 | " + clean_text(str(m.get('student_id',''))) + " |\n"
+          "| 班级 | " + clean_text(str(m.get('class',''))) + " |\n"
+          "| 成绩 | " + clean_text(str(m.get('score',''))) + " 分 |\n"
+          "| 类型 | " + tm.get(s.get('task_type',''),'') + " |\n")
+    return md, clean_text(s.get('question','')), clean_text(s.get('answer',''))
+
+
+ATK_MAP = {"基线模型 (Baseline)":"baseline","标签平滑 (e=0.02)":"smooth_0.02","标签平滑 (e=0.2)":"smooth_0.2",
+           "输出扰动 (s=0.01)":"perturbation_0.01","输出扰动 (s=0.015)":"perturbation_0.015","输出扰动 (s=0.02)":"perturbation_0.02"}
+
+
+def cb_attack(idx, src, target):
+    is_mem = src == "成员数据（训练集）"
+    pool = member_data if is_mem else non_member_data
+    idx = min(int(idx), len(pool)-1)
+    sample = pool[idx]
+    key = ATK_MAP.get(target, "baseline")
+    is_op = key.startswith("perturbation_")
+
+    if is_op:
+        sigma = float(key.split("_")[1])
+        fr = full_results.get('baseline', {})
+        lk = 'member_losses' if is_mem else 'non_member_losses'
+        base_loss = fr[lk][idx] if idx < len(fr.get(lk, [])) else float(np.random.normal(bl_m_mean if is_mem else bl_nm_mean, .02))
         np.random.seed(idx * 1000 + int(sigma * 1000))
         loss = base_loss + np.random.normal(0, sigma)
-        m_mean, nm_mean, m_std_v, nm_std_v = bl_m_mean, bl_nm_mean, bl_m_std, bl_nm_std
-        model_auc = perturb_results.get(model_key, {}).get('auc', 0)
-        display_label = "OP(s=" + str(sigma) + ")"
+        mm, nm, ms, ns = bl_m_mean, bl_nm_mean, bl_m_std, bl_nm_std
+        auc_v = perturb_results.get(key, {}).get('auc', 0)
+        lbl = f"OP(s={sigma})"
     else:
-        params = MODEL_PARAMS.get(model_key, MODEL_PARAMS["baseline"])
-        fr = full_results.get(model_key, full_results.get('baseline', {}))
-        losses_key = 'member_losses' if is_member else 'non_member_losses'
-        if idx < len(fr.get(losses_key, [])):
-            loss = fr[losses_key][idx]
-        else:
-            loss = float(np.random.normal(params['m_mean'] if is_member else params['nm_mean'], 0.02))
-        m_mean, nm_mean = params['m_mean'], params['nm_mean']
-        m_std_v, nm_std_v = params['m_std'], params['nm_std']
-        model_auc = mia_results.get(model_key, {}).get('auc', 0)
-        display_label = params['label']
-
-    threshold = (m_mean + nm_mean) / 2.0
-    pred_member = (loss < threshold)
-    attack_correct = (pred_member == is_member)
-    gauge_fig = make_loss_gauge(loss, m_mean, nm_mean, threshold, m_std_v, nm_std_v)
-
-    pl = "训练成员" if pred_member else "非训练成员"
-    pc = "🔴" if pred_member else "🟢"
-    al = "训练成员" if is_member else "非训练成员"
-    ac = "🔴" if is_member else "🟢"
-
-    if attack_correct and pred_member and is_member:
-        v = "⚠️ **攻击成功: 发生了隐私泄露**"; vd = "模型对该样本过于熟悉（Loss低于阈值），攻击者成功判定其为训练集数据。"
-    elif attack_correct:
-        v = "✅ **判断正确**"; vd = "攻击者的判定与真实身份一致。"
+        info = MODEL_INFO.get(key, MODEL_INFO['baseline'])
+        fr = full_results.get(key, full_results.get('baseline', {}))
+        lk = 'member_losses' if is_mem else 'non_member_losses'
+        loss = fr[lk][idx] if idx < len(fr.get(lk, [])) else float(np.random.normal(info['m_mean'] if is_mem else info['nm_mean'], .02))
+        mm, nm, ms, ns = info['m_mean'], info['nm_mean'], info['m_std'], info['nm_std']
+        auc_v = info['auc']
+        lbl = info['label']
+
+    thr = (mm + nm) / 2
+    pred = loss < thr
+    correct = pred == is_mem
+    gauge = fig_loss_gauge(loss, mm, nm, thr, ms, ns)
+
+    pl, pc = ("训练成员","🔴") if pred else ("非训练成员","🟢")
+    al, ac = ("训练成员","🔴") if is_mem else ("非训练成员","🟢")
+
+    if correct and pred and is_mem:
+        v = "⚠️ **攻击成功：隐私泄露**\n\n模型对该样本过于熟悉（Loss < 阈值），攻击者成功判定为训练数据。"
+    elif correct:
+        v = "✅ **判定正确**\n\n攻击者的判定与真实身份一致。"
     else:
-        v = "❌ **攻击失误**"; vd = "攻击者的判定与真实身份不符。"
-
-    result_md = (v + "\n\n" + vd + "\n\n"
-        "**当前攻击模型**: " + display_label + " (AUC=" + f"{model_auc:.4f}" + ")\n\n"
-        "| | 攻击者计算得出 | 系统真实身份 |\n|---|---|---|\n"
-        "| 判定 | " + pc + " " + pl + " | " + ac + " " + al + " |\n"
-        "| Loss | " + f"{loss:.4f}" + " | Threshold: " + f"{threshold:.4f}" + " |\n")
-    q_text = "**样本追踪号 [" + str(idx) + "] :**\n\n" + clean_text(sample.get('question',''))[:500]
-    return q_text, gauge_fig, result_md
-
-
-EVAL_MODEL_MAP = {
-    "基线模型 (Baseline)": "baseline",
-    "标签平滑模型 (e=0.02)": "smooth_0.02",
-    "标签平滑模型 (e=0.2)": "smooth_0.2",
-    "输出扰动 (s=0.01)": "baseline",
-    "输出扰动 (s=0.015)": "baseline",
-    "输出扰动 (s=0.02)": "baseline",
-}
+        v = "🛡️ **防御成功：攻击失误**\n\n攻击者的判定与真实身份不符，防御起到了作用。"
 
-EVAL_ACC_MAP = {
-    "基线模型 (Baseline)": bl_acc,
-    "标签平滑模型 (e=0.02)": s002_acc,
-    "标签平滑模型 (e=0.2)": s02_acc,
-    "输出扰动 (s=0.01)": bl_acc,
-    "输出扰动 (s=0.015)": bl_acc,
-    "输出扰动 (s=0.02)": bl_acc,
-}
+    res = (v + "\n\n**攻击目标**: " + lbl + "　|　**AUC**: " + f"{auc_v:.4f}" + "\n\n"
+           "| | 攻击者判定 | 真实身份 |\n|---|---|---|\n"
+           "| 身份 | " + pc + " " + pl + " | " + ac + " " + al + " |\n"
+           "| Loss | " + f"{loss:.4f}" + " | 阈值: " + f"{thr:.4f}" + " |\n")
+
+    qtxt = "**样本 #" + str(idx) + "**\n\n" + clean_text(sample.get('question', ''))[:500]
+    return qtxt, gauge, res
 
 
-def run_eval_demo(eval_model):
-    model_key = EVAL_MODEL_MAP.get(eval_model, "baseline")
-    overall_acc = EVAL_ACC_MAP.get(eval_model, bl_acc)
-    idx = np.random.randint(0, len(EVAL_QUESTIONS))
-    q = EVAL_QUESTIONS[idx]
-    is_correct = q.get(model_key, q.get('baseline', False))
-    icon = "✅" if is_correct else "❌"
-    result_md = (
-        "### 测试结果\n\n"
-        "**模型**: " + eval_model + " (总体准确率: " + f"{overall_acc:.1f}" + "%)\n\n"
-        "| 项目 | 内容 |\n|---|---|\n"
-        "| 题目编号 | #" + str(idx+1) + " / 300 |\n"
-        "| 题目类型 | " + q.get('type_cn', q['type']) + " |\n"
-        "| 题目 | " + q['question'] + " |\n"
-        "| 正确答案 | " + q['answer'] + " |\n"
-        "| 模型判定 | " + icon + " " + ("正确" if is_correct else "错误") + " |\n\n")
-    if eval_model.startswith("输出扰动"):
-        result_md += "> 输出扰动不改变模型参数，因此准确率与基线完全一致。\n"
-    return result_md
+EVAL_ACC = {"基线模型":bl_acc,"标签平滑 (e=0.02)":s002_acc,"标签平滑 (e=0.2)":s02_acc,
+            "输出扰动 (s=0.01)":bl_acc,"输出扰动 (s=0.015)":bl_acc,"输出扰动 (s=0.02)":bl_acc}
+EVAL_KEY = {"基线模型":"baseline","标签平滑 (e=0.02)":"smooth_0.02","标签平滑 (e=0.2)":"smooth_0.2",
+            "输出扰动 (s=0.01)":"baseline","输出扰动 (s=0.015)":"baseline","输出扰动 (s=0.02)":"baseline"}
+
+
+def cb_eval(model):
+    k = EVAL_KEY.get(model, "baseline")
+    acc = EVAL_ACC.get(model, bl_acc)
+    q = EVAL_POOL[np.random.randint(len(EVAL_POOL))]
+    ok = q.get(k, q.get('baseline', False))
+    ic = "✅ 正确" if ok else "❌ 错误"
+    note = "\n\n> 输出扰动不改变模型，准确率与基线一致。" if "扰动" in model else ""
+    return ("**" + model + "**　总体��确率: " + f"{acc:.1f}%" + "\n\n"
+            "| 项目 | 内容 |\n|---|---|\n"
+            "| 类型 | " + q['type_cn'] + " |\n"
+            "| 题目 | " + q['question'] + " |\n"
+            "| 正确答案 | " + q['answer'] + " |\n"
+            "| 判定 | " + ic + " |" + note)
 
 
 # ========================================
-# Interface
+# 界面
 # ========================================
 
 CSS = """
-body { background-color: #f0f4f8 !important; }
-.gradio-container { max-width: 1200px !important; margin: auto !important; font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, "PingFang SC", "Microsoft YaHei", sans-serif !important; }
-.tab-nav { border-bottom: 2px solid #e1e8f0 !important; margin-bottom: 20px !important; }
-.tab-nav button { font-size: 15px !important; padding: 14px 22px !important; font-weight: 500 !important; color: #64748b !important; border-radius: 8px 8px 0 0 !important; background: transparent !important; border: none !important; }
-.tab-nav button.selected { font-weight: 700 !important; color: #2563eb !important; border-bottom: 3px solid #2563eb !important; }
-.tabitem { background: #fff !important; border-radius: 12px !important; box-shadow: 0 4px 20px rgba(0,0,0,0.04) !important; padding: 30px !important; border: 1px solid #e2e8f0 !important; }
-.prose h1 { font-size: 2rem !important; color: #0f172a !important; font-weight: 800 !important; text-align: center !important; }
-.prose h2 { font-size: 1.35rem !important; color: #1e293b !important; margin-top: 1.5em !important; padding-bottom: 0.4em !important; border-bottom: 2px solid #f1f5f9 !important; font-weight: 700 !important; }
-.prose h3 { font-size: 1.1rem !important; color: #334155 !important; font-weight: 600 !important; }
-.prose table { width: 100% !important; border-collapse: separate !important; border-spacing: 0 !important; margin: 1.2em 0 !important; border-radius: 10px !important; overflow: hidden !important; box-shadow: 0 0 0 1px #e2e8f0, 0 4px 6px -1px rgba(0,0,0,0.05) !important; font-size: 0.9rem !important; }
-.prose th { background: #f8fafc !important; color: #475569 !important; font-weight: 600 !important; padding: 10px 14px !important; border-bottom: 2px solid #e2e8f0 !important; }
-.prose tr:nth-child(even) td { background: #f8fafc !important; }
-.prose td { padding: 9px 14px !important; color: #334155 !important; border-bottom: 1px solid #e2e8f0 !important; }
-.prose blockquote { border-left: 4px solid #3b82f6 !important; background: linear-gradient(to right,#eff6ff,#fff) !important; padding: 14px 18px !important; border-radius: 0 8px 8px 0 !important; color: #1e40af !important; }
-button.primary { background: linear-gradient(135deg,#3b82f6 0%,#2563eb 100%) !important; border: none !important; box-shadow: 0 4px 12px rgba(37,99,235,0.25) !important; font-weight: 600 !important; }
+:root { --blue: #2563eb; --slate: #334155; }
+body { background: #f8fafc !important; }
+.gradio-container { max-width: 1180px !important; margin: auto !important;
+    font-family: "Inter", -apple-system, "PingFang SC", "Microsoft YaHei", sans-serif !important; }
+
+/* Tab */
+.tab-nav { border-bottom: 2px solid #e2e8f0 !important; gap: 4px !important; }
+.tab-nav button { font-size: 14px !important; padding: 12px 20px !important; font-weight: 500 !important;
+    color: #64748b !important; border: none !important; background: transparent !important;
+    border-radius: 6px 6px 0 0 !important; transition: .2s !important; }
+.tab-nav button:hover { color: var(--blue) !important; background: #eff6ff !important; }
+.tab-nav button.selected { color: var(--blue) !important; font-weight: 700 !important;
+    border-bottom: 2.5px solid var(--blue) !important; background: #eff6ff !important; }
+
+.tabitem { background: #fff !important; border-radius: 0 0 10px 10px !important;
+    box-shadow: 0 1px 3px rgba(0,0,0,.04) !important; padding: 28px !important;
+    border: 1px solid #e2e8f0 !important; border-top: none !important; }
+
+/* Typography */
+.prose h1 { font-size: 1.75rem !important; color: #0f172a !important; font-weight: 800 !important;
+    text-align: center !important; letter-spacing: -.02em !important; margin-bottom: .3em !important; }
+.prose h2 { font-size: 1.2rem !important; color: #1e293b !important; font-weight: 700 !important;
+    margin-top: 1.4em !important; padding-bottom: .3em !important; border-bottom: 1.5px solid #f1f5f9 !important; }
+.prose h3 { font-size: 1.02rem !important; color: var(--slate) !important; font-weight: 600 !important; }
+
+/* Table */
+.prose table { width: 100% !important; border-collapse: separate !important; border-spacing: 0 !important;
+    border-radius: 8px !important; overflow: hidden !important; margin: 1em 0 !important;
+    box-shadow: 0 0 0 1px #e2e8f0 !important; font-size: .88rem !important; }
+.prose th { background: #f8fafc !important; color: #475569 !important; font-weight: 600 !important;
+    padding: 9px 12px !important; border-bottom: 1.5px solid #e2e8f0 !important; font-size: .82rem !important; }
+.prose td { padding: 8px 12px !important; color: var(--slate) !important; border-bottom: 1px solid #f1f5f9 !important; }
+.prose tr:last-child td { border-bottom: none !important; }
+
+/* Blockquote */
+.prose blockquote { border-left: 3px solid var(--blue) !important; background: #f0f7ff !important;
+    padding: 10px 14px !important; border-radius: 0 6px 6px 0 !important; color: #1e40af !important;
+    font-size: .9rem !important; margin: 1em 0 !important; }
+
+/* Button */
+button.primary { background: var(--blue) !important; border: none !important;
+    box-shadow: 0 2px 8px rgba(37,99,235,.2) !important; font-weight: 600 !important; border-radius: 8px !important; }
+
 footer { display: none !important; }
 """
 
-with gr.Blocks(title="教育大模型隐私攻防", theme=gr.themes.Soft(primary_hue="blue", secondary_hue="sky", neutral_hue="slate"), css=CSS) as demo:
+with gr.Blocks(title="MIA攻防研究", theme=gr.themes.Soft(primary_hue="blue", neutral_hue="slate"), css=CSS) as demo:
 
-    gr.Markdown("# 教育大模型中的成员推理攻击及其防御研究\n\n> 探究教育场景下大语言模型的隐私泄露风险，验证标签平滑与输出扰动两种防御策略的有效性。\n")
+    gr.Markdown("# 教育大模型中的成员推理攻击及其防御研究\n"
+                "> 基于 " + model_name + " 微调的数学辅导模型，验证MIA风险与两类防御策略的有效性")
 
-    with gr.Tab("项目概览"):
+    # ────────────── Tab 1 ──────────────
+    with gr.Tab("实验总览"):
         gr.Markdown(
-            "## 研究背景\n\n大语言模型在教育领域广泛应用，训练过程不可避免接触学生敏感数据。**成员推理攻击 (MIA)** 能判断数据是否参与训练，构成隐私威胁。\n\n---\n\n"
-            "## 实验设计\n\n"
+            "## 研究问题\n\n"
+            "如果用包含学生隐私的数据训练教育AI，攻击者能否推断出哪些学生的数据被使用？\n\n---\n\n"
+            "## 核心指标\n\n"
+            "| 指标 | 基线 | LS(e=0.02) | LS(e=0.2) | OP(s=0.01) | OP(s=0.015) | OP(s=0.02) |\n"
+            "|------|------|-----------|----------|-----------|------------|----------|\n"
+            "| AUC | " + f"{bl_auc:.4f}" + " | " + f"{s002_auc:.4f}" + " | " + f"{s02_auc:.4f}" + " | " + f"{op001_auc:.4f}" + " | " + f"{op0015_auc:.4f}" + " | " + f"{op002_auc:.4f}" + " |\n"
+            "| 准确率 | " + f"{bl_acc:.1f}%" + " | " + f"{s002_acc:.1f}%" + " | " + f"{s02_acc:.1f}%" + " | " + f"{bl_acc:.1f}%" + " | " + f"{bl_acc:.1f}%" + " | " + f"{bl_acc:.1f}%" + " |\n\n"
+            "> AUC越接近0.5，防御越有效。准确率越高，模型效用越好。\n\n---\n\n"
+            "## 实验流程\n\n"
             "| 阶段 | 内容 | 方法 |\n|------|------|------|\n"
-            "| 1. 数据准备 | 2000条小学数学辅导对话 | 模板化生成，含隐私字段 |\n"
-            "| 2. 基线模型训练 | Qwen2.5-Math-1.5B + LoRA | 标准微调，无防御 |\n"
-            "| 3. 标签平滑模型训练 | 两组平滑系数 | e=0.02 与 e=0.2 分别训练 |\n"
-            "| 4. MIA攻击测试 | 全部模型及策略 | 三模型Loss攻击 + 三组输出扰动 |\n"
-            "| 5. 效用评估 | 300道数学测试题 | 三模型 + 三组扰动分别测试 |\n"
-            "| 6. 综合分析 | 隐私-效用权衡 | 散点图 + 定量对比 |\n\n---\n\n"
-            "## 实验配置\n\n| 项目 | 值 |\n|------|-----|\n"
-            "| 基座模型 | " + model_name_str + " |\n"
-            "| 微调 | LoRA (r=8, alpha=16) |\n| 训练轮数 | 10 epochs |\n"
-            "| 数据量 | " + data_size_str + " 条 |\n| 模型数 | 3个 |\n")
-
-    with gr.Tab("数据展示"):
-        gr.Markdown("## 数据集概况\n\n"
-            "- **成员数据** (1000条): 用于模型训练，模型会\"记住\"这些数据\n"
-            "- **非成员数据** (1000条): 不参与训练，作为攻击对照组\n"
-            "- 两组数据**格式完全相同**（都含隐私字段），这是MIA实验的标准设置——攻击者无法从数据格式区分成员与非成员\n\n"
-            "### 任务类型分布\n\n"
-            "| 类型 | 数量 | 占比 |\n|------|------|------|\n"
-            "| 基础计算 | 800 | 40% |\n| 应用题 | 600 | 30% |\n| 概念问答 | 400 | 20% |\n| 错题订正 | 200 | 10% |\n")
+            "| 1. 数据准备 | 2000条数学辅导对话 | 模板化生成，含隐私字段 |\n"
+            "| 2. 基线训练 | Qwen2.5-Math + LoRA | 标准微调，无防御 |\n"
+            "| 3. 防御训练 | 标签平滑 e=0.02 / 0.2 | 两组参数分别训练 |\n"
+            "| 4. 攻击测试 | 3个模型 + 3组输出扰动 | Loss阈值判定，AUC评估 |\n"
+            "| 5. 效用评估 | 300道数学题 | 6种配置分别测试 |\n"
+            "| 6. 综合分析 | 隐私-效用权衡 | 定量对比 |\n\n"
+            "## 实验配置\n\n"
+            "| 项目 | 值 |\n|---|---|\n"
+            "| 模型 | " + model_name + " |\n"
+            "| 微调 | LoRA (r=8, alpha=16) |\n"
+            "| 训练 | 10 epochs |\n"
+            "| 数据 | 成员1000条 + 非成员1000条 |\n")
+
+    # ────────────── Tab 2 ──────────────
+    with gr.Tab("数据与模型"):
+        gr.Markdown(
+            "## 数据集\n\n"
+            "- **成员数据** (1000条)：用于模型训练，模型会\"记住\"这些数据\n"
+            "- **非成员数据** (1000条)：不参与训练，作为攻击的对照组\n"
+            "- 两组数据格式完全相同（均��隐私字段），这是MIA实验的标准设置\n\n"
+            "| 任务类型 | 数量 | 占比 |\n|---|---|---|\n"
+            "| 基础计算 | 800 | 40% |\n| 应用题 | 600 | 30% |\n| 概念问答 | 400 | 20% |\n| 错题订正 | 200 | 10% |\n\n"
+            "### 数据样例\n\n选择数据池并随机提取一条样本，查看其包含的隐私信息和对话内容。")
         with gr.Row():
-            with gr.Column():
-                data_sel = gr.Radio(["成员数据（训练集）","非成员数据（测试集）"], value="成员数据（训练集）", label="选择数据池")
-                sample_btn = gr.Button("随机提取", variant="primary")
-                sample_info = gr.Markdown()
-            with gr.Column():
-                sample_q = gr.Textbox(label="学生提问 (Prompt)", lines=5, interactive=False)
-                sample_a = gr.Textbox(label="模型回答 (Ground Truth)", lines=5, interactive=False)
-        sample_btn.click(show_random_sample, [data_sel], [sample_info, sample_q, sample_a])
-
-    with gr.Tab("MIA攻击演示"):
-        gr.Markdown("## 发起成员推理攻击\n\n选择攻击目标和数据来源，系统将计算Loss并判定。\n")
+            with gr.Column(scale=2):
+                d_src = gr.Radio(["成员数据（训练集）","非成员数据（测试集）"], value="成员数据（训练集）", label="数据来源")
+                d_btn = gr.Button("随机提取样本", variant="primary")
+                d_meta = gr.Markdown()
+            with gr.Column(scale=3):
+                d_q = gr.Textbox(label="学生提问", lines=4, interactive=False)
+                d_a = gr.Textbox(label="标准回答", lines=4, interactive=False)
+        d_btn.click(cb_sample, [d_src], [d_meta, d_q, d_a])
+
+    # ────────────── Tab 3 ──────────────
+    with gr.Tab("攻击与防御验证"):
+        gr.Markdown("## 交互式MIA攻击演示\n\n"
+                    "通过对照实验验证攻击的有效性和防御策略的效果。\n\n"
+                    "**建议操作顺序**: ① 基线+成员 → ② 基线+非成员 → ③ 标签平滑+成员 → ④ 输出扰动+成员")
         with gr.Row():
-            with gr.Column():
-                atk_model = gr.Radio(["基线模型 (Baseline)","标签平滑模型 (e=0.02)","标签平滑模型 (e=0.2)",
-                    "输出扰动 (s=0.01)","输出扰动 (s=0.015)","输出扰动 (s=0.02)"], value="基线模型 (Baseline)", label="选择攻击目标")
-                atk_type = gr.Radio(["成员数据（训练集）","非成员数据（测试集）"], value="成员数据（训练集）", label="数据来源")
-                atk_idx = gr.Slider(0, 999, step=1, value=0, label="样本ID (0-999)")
-                atk_btn = gr.Button("执行成员推理攻击", variant="primary", size="lg")
-                atk_question = gr.Markdown()
-            with gr.Column():
-                gr.Markdown("**攻击侦测控制台**")
-                atk_gauge = gr.Plot(label="Loss分布雷达")
-                atk_result = gr.Markdown()
-        atk_btn.click(run_mia_demo, [atk_idx, atk_type, atk_model], [atk_question, atk_gauge, atk_result])
-
-    with gr.Tab("防御对比"):
-        gr.Markdown("## 防御策略效果对比\n\n"
-            "| 策略 | 类型 | 原理 | 实验优势 | 实验局限 |\n|------|------|------|---------|--------|\n"
-            "| 标签平滑 | 训练期 | 软化标签抑制过度记忆 | AUC降至" + f"{s002_auc:.4f}" + "(e=0.02) | 需重新训练 |\n"
-            "| 输出扰动 | 推理期 | Loss加高斯噪声 | AUC降至" + f"{op002_auc:.4f}" + "(s=0.02)，零效用损失 | 仅遮蔽统计信号 |\n")
-        gr.Markdown("### AUC对比"); gr.Plot(value=make_auc_bar())
-        gr.Markdown("### Loss分布 - 三个模型"); gr.Plot(value=make_loss_distribution())
-        gr.Markdown("### Loss分布 - 输出扰动效果"); gr.Plot(value=make_perturb_loss_distribution())
-        tbl = "### 完整结果\n\n| 策略 | 类型 | AUC | 准确率 | AUC变化 |\n|------|------|-----|--------|--------|\n"
-        for k, n, cat in [('baseline','基线','--'),('smooth_0.02','LS(e=0.02)','训练期'),('smooth_0.2','LS(e=0.2)','训练期')]:
-            if k in mia_results:
-                a=mia_results[k]['auc']; acc=utility_results.get(k,{}).get('accuracy',0)*100
-                d = "--" if k=='baseline' else f"{a-bl_auc:+.4f}"
-                tbl += "| "+n+" | "+cat+" | "+f"{a:.4f}"+" | "+f"{acc:.1f}"+"%"+" | "+d+" |\n"
-        for k, n in [('perturbation_0.01','OP(s=0.01)'),('perturbation_0.015','OP(s=0.015)'),('perturbation_0.02','OP(s=0.02)')]:
-            if k in perturb_results:
-                a=perturb_results[k]['auc']
-                tbl += "| "+n+" | 推理期 | "+f"{a:.4f}"+" | "+f"{bl_acc:.1f}"+"% (不变) | "+f"{a-bl_auc:+.4f}"+" |\n"
-        gr.Markdown(tbl)
-
-    with gr.Tab("防御详解"):
+            with gr.Column(scale=2):
+                a_target = gr.Radio(["基线模型 (Baseline)","标签平滑 (e=0.02)","标签平滑 (e=0.2)",
+                    "输出扰动 (s=0.01)","输出扰动 (s=0.015)","输出扰动 (s=0.02)"],
+                    value="基线模型 (Baseline)", label="攻击目标")
+                a_src = gr.Radio(["成员数据（训练集）","非成员数据（测试集）"], value="成员数据（训练集）", label="数据来源")
+                a_idx = gr.Slider(0, 999, step=1, value=12, label="样本 ID")
+                a_btn = gr.Button("执行攻击", variant="primary", size="lg")
+                a_qtxt = gr.Markdown()
+            with gr.Column(scale=3):
+                a_gauge = gr.Plot(label="Loss位置")
+                a_res = gr.Markdown()
+        a_btn.click(cb_attack, [a_idx, a_src, a_target], [a_qtxt, a_gauge, a_res])
+
+    # ────────────── Tab 4 ──────────────
+    with gr.Tab("实验结果分析"):
+        gr.Markdown("## 攻击与防御效果\n")
+        gr.Markdown("### MIA攻击AUC对比")
+        gr.Plot(value=fig_auc_bar())
+        gr.Markdown("### Loss分布 — 三个模型")
+        gr.Plot(value=fig_loss_dist())
+        gr.Markdown("### Loss分布 — 输出扰动效果")
+        gr.Plot(value=fig_perturb_dist())
+
         gr.Markdown(
-            "## 一、标签平滑 (Label Smoothing)\n\n**类型**: 训练期防御\n\n"
-            "将训练标签从硬标签转换为软标签，降低过拟合。\n\n"
-            "**公式**: y_smooth = (1 - e) * y_onehot + e / V\n\n"
-            "其中 e 为平滑系数，V 为词汇表大小。\n\n"
-            "| 参数 | AUC | 准确率 | 分析 |\n|------|-----|--------|------|\n"
-            "| 基线 (e=0) | " + f"{bl_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% | 无防御 |\n"
-            "| e=0.02 | " + f"{s002_auc:.4f}" + " | " + f"{s002_acc:.1f}" + "% | 温和平滑 |\n"
-            "| e=0.2 | " + f"{s02_auc:.4f}" + " | " + f"{s02_acc:.1f}" + "% | 强力平滑 |\n\n---\n\n"
-            "## 二、输出扰动 (Output Perturbation)\n\n**类型**: 推理期防御\n\n"
-            "在推理阶段对Loss注入高斯噪声。\n\n"
-            "**公式**: L_perturbed = L_original + N(0, s^2)\n\n"
-            "| 参数 | AUC | AUC降幅 | 准确率 |\n|------|-----|---------|--------|\n"
-            "| 基线 | " + f"{bl_auc:.4f}" + " | -- | " + f"{bl_acc:.1f}" + "% |\n"
-            "| s=0.01 | " + f"{op001_auc:.4f}" + " | " + f"{bl_auc-op001_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% |\n"
-            "| s=0.015 | " + f"{op0015_auc:.4f}" + " | " + f"{bl_auc-op0015_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% |\n"
-            "| s=0.02 | " + f"{op002_auc:.4f}" + " | " + f"{bl_auc-op002_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% |\n\n---\n\n"
-            "## 三、综合对比\n\n| 维度 | 标签平滑 | 输出扰动 |\n|------|---------|----------|\n"
-            "| 作用阶段 | 训练期 | 推理期 |\n| 需要重训 | 是 | 否 |\n| 效用影响 | 取决于系数 | 无 |\n| 防御原理 | 降低记忆 | 遮蔽信号 |\n| 部署难度 | 训练介入 | 即插即用 |\n")
-
-    with gr.Tab("效用评估"):
-        gr.Markdown("## 效用评估\n\n> 从300道测试题中随机抽取，展示模型的实际作答情况。\n")
+            "### 完整结果\n\n"
+            "| 策略 | 类型 | AUC | 准确率 | AUC变化 |\n|---|---|---|---|---|\n"
+            "| 基线 | — | " + f"{bl_auc:.4f}" + " | " + f"{bl_acc:.1f}%" + " | — |\n"
+            "| LS(e=0.02) | 训练期 | " + f"{s002_auc:.4f}" + " | " + f"{s002_acc:.1f}%" + " | " + f"{s002_auc-bl_auc:+.4f}" + " |\n"
+            "| LS(e=0.2) | 训练期 | " + f"{s02_auc:.4f}" + " | " + f"{s02_acc:.1f}%" + " | " + f"{s02_auc-bl_auc:+.4f}" + " |\n"
+            "| OP(s=0.01) | 推理期 | " + f"{op001_auc:.4f}" + " | " + f"{bl_acc:.1f}%" + " | " + f"{op001_auc-bl_auc:+.4f}" + " |\n"
+            "| OP(s=0.015) | 推理期 | " + f"{op0015_auc:.4f}" + " | " + f"{bl_acc:.1f}%" + " | " + f"{op0015_auc-bl_auc:+.4f}" + " |\n"
+            "| OP(s=0.02) | 推理期 | " + f"{op002_auc:.4f}" + " | " + f"{bl_acc:.1f}%" + " | " + f"{op002_auc-bl_auc:+.4f}" + " |\n")
+
+        gr.Markdown("---\n## 效用评估\n")
         with gr.Row():
-            with gr.Column():
-                gr.Markdown("### 准确率对比"); gr.Plot(value=make_accuracy_bar())
-            with gr.Column():
-                gr.Markdown("### 隐私-效用权衡"); gr.Plot(value=make_tradeoff())
-        gr.Markdown("### 在线效用测试")
+            with gr.Column(): gr.Plot(value=fig_acc_bar())
+            with gr.Column(): gr.Plot(value=fig_tradeoff())
+
+        gr.Markdown("### 在线效用测试\n\n随机抽取测试题，查看不同模型的作答表现。")
         with gr.Row():
-            with gr.Column():
-                eval_model = gr.Radio(["基线模型 (Baseline)","标签平滑模型 (e=0.02)","标签平滑模型 (e=0.2)",
-                    "输出扰动 (s=0.01)","输出扰动 (s=0.015)","输出扰动 (s=0.02)"], value="基线模型 (Baseline)", label="选择模型/策略")
-                eval_btn = gr.Button("随机抽题测试", variant="primary")
-            with gr.Column():
-                eval_result = gr.Markdown()
-        eval_btn.click(run_eval_demo, [eval_model], [eval_result])
-
-    with gr.Tab("实验结果可视化"):
-        gr.Markdown("## 实验核心图表")
-        for fn, cap in [("fig1_loss_distribution_comparison.png","图1: 成员与非成员Loss分布对比"),
-                         ("fig2_privacy_utility_tradeoff_fixed.png","图2: 隐私风险与模型效用权衡"),
-                         ("fig3_defense_comparison_bar.png","图3: 各防御策略AUC对比")]:
-            p = os.path.join(BASE_DIR,"figures",fn)
+            with gr.Column(scale=1):
+                e_model = gr.Radio(["基线模型","标签平滑 (e=0.02)","标签平滑 (e=0.2)",
+                    "输出扰动 (s=0.01)","输出扰动 (s=0.015)","输出扰动 (s=0.02)"], value="基线模型", label="模型")
+                e_btn = gr.Button("随机抽题", variant="primary")
+            with gr.Column(scale=2):
+                e_res = gr.Markdown()
+        e_btn.click(cb_eval, [e_model], [e_res])
+
+        gr.Markdown("---\n### 防御机制对比\n\n"
+            "| 维度 | 标签平滑 | 输出扰动 |\n|---|---|---|\n"
+            "| 阶段 | 训练期 | 推理期 |\n"
+            "| 原理 | 软化标签，降低记忆 | Loss加噪声，模糊信号 |\n"
+            "| 需重训 | 是 | 否 |\n"
+            "| 效用影响 | 取决于参数 | 无 |\n"
+            "| 部署 | 训练时介入 | 即插即用 |\n\n"
+            "**标签平滑公式**: y_smooth = (1-e) * y_onehot + e/V\n\n"
+            "**输出扰动公式**: L_perturbed = L_original + N(0, s^2)\n")
+
+        for fn, cap in [("fig1_loss_distribution_comparison.png","Loss分布对比"),
+                         ("fig2_privacy_utility_tradeoff_fixed.png","隐私-效用权衡"),
+                         ("fig3_defense_comparison_bar.png","防御策略AUC对比")]:
+            p = os.path.join(BASE_DIR, "figures", fn)
             if os.path.exists(p):
-                gr.Markdown("### "+cap); gr.Image(value=p, show_label=False, height=450); gr.Markdown("---")
+                gr.Markdown("### " + cap); gr.Image(value=p, show_label=False, height=420)
 
+    # ────────────── Tab 5 ──────────────
     with gr.Tab("研究结论"):
         gr.Markdown(
-            "## 研究结论\n\n---\n\n"
-            "### 一、教育大模型面临显著的MIA风险\n\n"
-            "基线模型 AUC = **" + f"{bl_auc:.4f}" + "**，成员平均Loss (" + f"{bl_m_mean:.4f}" + ") 低于非成员 (" + f"{bl_nm_mean:.4f}" + ")，模型对训练数据存在可被利用的记忆效应。\n\n---\n\n"
-            "### 二、标签平滑的有效性与局限性\n\n"
-            "| 参数 | AUC | 准确率 | 分析 |\n|------|-----|--------|------|\n"
-            "| 基线 (e=0) | " + f"{bl_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% | 无防御 |\n"
-            "| e=0.02 | " + f"{s002_auc:.4f}" + " | " + f"{s002_acc:.1f}" + "% | 正则化提升泛化 |\n"
-            "| e=0.2 | " + f"{s02_auc:.4f}" + " | " + f"{s02_acc:.1f}" + "% | 防御更强 |\n\n"
-            "e=0.02在隐私保护与效用保持间取得较好平衡。\n\n---\n\n"
-            "### 三、输出扰动的独特优势\n\n"
-            "| 参数 | AUC | AUC降幅 | 准确率 |\n|------|-----|---------|--------|\n"
-            "| s=0.01 | " + f"{op001_auc:.4f}" + " | " + f"{bl_auc-op001_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% |\n"
-            "| s=0.015 | " + f"{op0015_auc:.4f}" + " | " + f"{bl_auc-op0015_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% |\n"
-            "| s=0.02 | " + f"{op002_auc:.4f}" + " | " + f"{bl_auc-op002_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% |\n\n"
-            "零效用损失，适合已部署系统加固。\n\n---\n\n"
+            "## 核心发现\n\n---\n\n"
+            "### 一、教育大模型存在可量化的MIA风险\n\n"
+            "基线模型 AUC = **" + f"{bl_auc:.4f}" + "**，显著高于随机基准0.5。成员平均Loss ("
+            + f"{bl_m_mean:.4f}" + ") 低于非成员 (" + f"{bl_nm_mean:.4f}" + ")，模型对训练数据存在可利用的记忆效应。\n\n---\n\n"
+            "### 二、标签平滑（训练期防御）\n\n"
+            "| 参数 | AUC | 准确率 | 分析 |\n|---|---|---|---|\n"
+            "| 基线 | " + f"{bl_auc:.4f}" + " | " + f"{bl_acc:.1f}%" + " | 无防御 |\n"
+            "| e=0.02 | " + f"{s002_auc:.4f}" + " | " + f"{s002_acc:.1f}%" + " | 正则化提升泛化 |\n"
+            "| e=0.2 | " + f"{s02_auc:.4f}" + " | " + f"{s02_acc:.1f}%" + " | 防御增强 |\n\n---\n\n"
+            "### 三、输出扰动（推理期防御）\n\n"
+            "| 参数 | AUC | AUC降幅 | 准确率 |\n|---|---|---|---|\n"
+            "| s=0.01 | " + f"{op001_auc:.4f}" + " | " + f"{bl_auc-op001_auc:.4f}" + " | " + f"{bl_acc:.1f}%" + " |\n"
+            "| s=0.015 | " + f"{op0015_auc:.4f}" + " | " + f"{bl_auc-op0015_auc:.4f}" + " | " + f"{bl_acc:.1f}%" + " |\n"
+            "| s=0.02 | " + f"{op002_auc:.4f}" + " | " + f"{bl_auc-op002_auc:.4f}" + " | " + f"{bl_acc:.1f}%" + " |\n\n"
+            "零效用损失，适合已部署系统的后期加固。\n\n---\n\n"
             "### 四、隐私-效用权衡\n\n"
-            "| 策略 | AUC | 准确率 | AUC变化 | 效用变化 |\n|------|-----|--------|--------|--------|\n"
-            "| 基线 | " + f"{bl_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% | -- | -- |\n"
-            "| LS e=0.02 | " + f"{s002_auc:.4f}" + " | " + f"{s002_acc:.1f}" + "% | " + f"{s002_auc-bl_auc:+.4f}" + " | " + f"{s002_acc-bl_acc:+.1f}" + "pp |\n"
-            "| LS e=0.2 | " + f"{s02_auc:.4f}" + " | " + f"{s02_acc:.1f}" + "% | " + f"{s02_auc-bl_auc:+.4f}" + " | " + f"{s02_acc-bl_acc:+.1f}" + "pp |\n"
-            "| OP s=0.01 | " + f"{op001_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% | " + f"{op001_auc-bl_auc:+.4f}" + " | 0 |\n"
-            "| OP s=0.015 | " + f"{op0015_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% | " + f"{op0015_auc-bl_auc:+.4f}" + " | 0 |\n"
-            "| OP s=0.02 | " + f"{op002_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% | " + f"{op002_auc-bl_auc:+.4f}" + " | 0 |\n\n"
-            "两类策略机制互补，可根据场景灵活选择或组合。\n")
-
-    gr.Markdown("---\n\n<center>教育大模型中的成员推理攻击及其防御思路研究</center>\n")
+            "| 策略 | AUC | 准确率 | 隐私 | 效用 |\n|---|---|---|---|---|\n"
+            "| 基线 | " + f"{bl_auc:.4f}" + " | " + f"{bl_acc:.1f}%" + " | 风险最高 | 基准 |\n"
+            "| LS(e=0.02) | " + f"{s002_auc:.4f}" + " | " + f"{s002_acc:.1f}%" + " | 风险降低 | 提升 |\n"
+            "| LS(e=0.2) | " + f"{s02_auc:.4f}" + " | " + f"{s02_acc:.1f}%" + " | 显著降低 | 可接受 |\n"
+            "| OP(s=0.02) | " + f"{op002_auc:.4f}" + " | " + f"{bl_acc:.1f}%" + " | 显著降低 | 不变 |\n\n"
+            "两类策略机制互补：标签平滑从训练阶段降低记忆，输出扰动从推理阶段遮蔽信号。\n")
+
+    gr.Markdown("<center style='color:#94a3b8;font-size:.85rem;margin-top:1em'>教育大模型成员推理攻击及其防御研究</center>")
 
 demo.launch()