Update app.py

app.py

@@ -65,13 +65,58 @@ MODEL_PARAMS = {
     "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)"},
 }
 
+# 预生成效用测试题库（模拟300道题的结果）
+EVAL_QUESTIONS = []
+eval_types = ['calculation'] * 120 + ['word_problem'] * 90 + ['concept'] * 60 + ['error_correction'] * 30
+np.random.seed(777)
+for i in range(300):
+    t = eval_types[i]
+    if t == 'calculation':
+        a, b = np.random.randint(10, 500), np.random.randint(10, 500)
+        ops = ['+', '-', 'x']
+        op = ops[i % 3]
+        if op == '+':
+            q = f"{a} + {b} = ?"
+            ans = str(a + b)
+        elif op == '-':
+            q = f"{a} - {b} = ?"
+            ans = str(a - b)
+        else:
+            q = f"{a} x {b} = ?"
+            ans = str(a * b)
+    elif t == 'word_problem':
+        a, b = np.random.randint(5, 100), np.random.randint(3, 50)
+        templates = [f"There are {a} apples, {b} are taken. How many left?",
+                     f"Each group has {a} people, {b} groups. Total?",
+                     f"A has {a} books, B has {b} more. B has?"]
+        q = templates[i % 3]
+        ans = str(a - b) if 'taken' in q else str(a * b) if 'group' in q else str(a + b)
+    elif t == 'concept':
+        concepts = ["area", "perimeter", "fraction", "decimal", "average"]
+        c = concepts[i % 5]
+        q = f"What is {c}?"
+        ans = f"Definition of {c}"
+    else:
+        a, b = np.random.randint(10, 99), np.random.randint(10, 99)
+        correct = a + b
+        wrong = correct + np.random.choice([-1, 1])
+        q = f"Student says {a}+{b}={wrong}. Correct?"
+        ans = str(correct)
+    # Simulate correctness per model
+    bl_correct = np.random.random() < (bl_acc / 100)
+    s002_correct = np.random.random() < (s002_acc / 100)
+    s02_correct = np.random.random() < (s02_acc / 100)
+    EVAL_QUESTIONS.append({
+        'question': q, 'answer': ans, 'type': t,
+        'baseline': bl_correct, 'smooth_0.02': s002_correct, 'smooth_0.2': s02_correct
+    })
+
 
 # ========================================
 # Charts
 # ========================================
 
 def make_loss_distribution():
-    """3 model Loss distributions - larger size"""
     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:
@@ -79,189 +124,126 @@ def make_loss_distribution():
             items.append((k, t + "\nAUC=" + f"{auc:.4f}"))
     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 * n, 5.5))
-    if n == 1:
-        axes = [axes]
+        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))
+    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']
+        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, loc='upper right')
-        ax.tick_params(labelsize=10)
+        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)
-        ax.spines['top'].set_visible(False)
-        ax.spines['right'].set_visible(False)
-    plt.suptitle('Model Loss Distribution: Member vs Non-Member', fontsize=15, fontweight='bold', y=1.02)
+        ax.spines['top'].set_visible(False); ax.spines['right'].set_visible(False)
     plt.tight_layout()
     return fig
 
 
 def make_perturb_loss_distribution():
-    """Output perturbation effect on baseline 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'])
-    sigmas = [0.01, 0.015, 0.02]
-    fig, axes = plt.subplots(1, 3, figsize=(18, 5.5))
-    for ax, sigma in zip(axes, sigmas):
+        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))
-        all_vals = np.concatenate([m_pert, nm_pert])
-        bins = np.linspace(all_vals.min(), all_vals.max(), 30)
-        ax.hist(m_pert, bins=bins, alpha=0.55, color='#5B8FF9', label='Member (perturbed)', density=True)
-        ax.hist(nm_pert, bins=bins, alpha=0.55, color='#E86452', label='Non-Member (perturbed)', density=True)
+        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, loc='upper right')
-        ax.tick_params(labelsize=10)
+        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.spines['top'].set_visible(False)
-        ax.spines['right'].set_visible(False)
-    plt.suptitle('Output Perturbation: Loss Distribution After Adding Noise', fontsize=15, fontweight='bold', y=1.02)
+        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, name, 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(name); aucs.append(mia_results[k]['auc']); colors.append(c)
-    for k, name, 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(name); aucs.append(perturb_results[k]['auc']); colors.append(c)
+    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.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 if aucs else 0.7)
-    ax.legend(fontsize=10)
-    ax.grid(axis='y', linestyle='--', alpha=0.3)
-    ax.spines['top'].set_visible(False)
-    ax.spines['right'].set_visible(False)
-    plt.xticks(fontsize=11)
-    plt.tight_layout()
+    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.spines['top'].set_visible(False); ax.spines['right'].set_visible(False)
+    plt.xticks(fontsize=11); plt.tight_layout()
     return fig
 
 
 def make_tradeoff():
     fig, ax = plt.subplots(figsize=(10, 7))
     pts = []
-    for k, name, 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)]:
+    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': name, '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, name, 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': name, 'a': perturb_results[k]['auc'], 'c': ba, 'm': mk, 'co': c, 's': sz})
+            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.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 (Privacy Risk)', fontsize=12, fontweight='bold')
+    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, fancybox=True)
-    ax.grid(True, alpha=0.2)
-    ax.spines['top'].set_visible(False)
-    ax.spines['right'].set_visible(False)
-    plt.tight_layout()
-    return fig
+    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)
+    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, name, 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(name); accs.append(utility_results[k]['accuracy']*100); colors.append(c)
-    bp = utility_results.get('baseline', {}).get('accuracy', 0)*100
-    for k, name, 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(name); accs.append(bp); colors.append(c)
+    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
+        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)
+    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, linestyle='-', 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=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.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-Member\n({nm_mean:.4f})', ha='center', va='top',
-            fontsize=8, color='#E86452', transform=ax.get_xaxis_transform())
-    in_member = loss_val < threshold
-    mc = '#5B8FF9' if in_member 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(),
+    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))
-    mc_x = (x_min + threshold) / 2
-    nmc_x = (threshold + x_max) / 2
-    ax.text(mc_x, 0.5, 'Member Zone', ha='center', va='center', fontsize=11,
-            color='#5B8FF9', fontweight='bold', alpha=0.5, transform=ax.get_xaxis_transform())
-    ax.text(nmc_x, 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
+    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
 
 
 # ========================================
@@ -272,16 +254,14 @@ 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', ''))
+    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 = {
@@ -295,99 +275,114 @@ MODEL_CHOICE_MAP = {
 
 
 def run_mia_demo(sample_index, data_type, model_choice):
-    is_member = (data_type == "成员数据��训练集）")
+    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)
+    idx = min(int(sample_index), len(data)-1)
     sample = data[idx]
-
     model_key = MODEL_CHOICE_MAP.get(model_choice, "baseline")
-
-    # Determine which Loss data to use
     is_perturb = model_key.startswith("perturbation_")
+
     if is_perturb:
-        # Output perturbation: baseline loss + noise
         sigma = float(model_key.split("_")[1])
         base_fr = full_results.get('baseline', {})
-        if is_member and idx < len(base_fr.get('member_losses', [])):
-            base_loss = base_fr['member_losses'][idx]
-        elif not is_member and idx < len(base_fr.get('non_member_losses', [])):
-            base_loss = base_fr['non_member_losses'][idx]
+        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))
         np.random.seed(idx * 1000 + int(sigma * 1000))
         loss = base_loss + np.random.normal(0, sigma)
-        m_mean = bl_m_mean
-        nm_mean = bl_nm_mean
-        m_std_v = bl_m_std
-        nm_std_v = bl_nm_std
+        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) + ")"
     else:
         params = MODEL_PARAMS.get(model_key, MODEL_PARAMS["baseline"])
         fr = full_results.get(model_key, full_results.get('baseline', {}))
-        if is_member and idx < len(fr.get('member_losses', [])):
-            loss = fr['member_losses'][idx]
-        elif not is_member and idx < len(fr.get('non_member_losses', [])):
-            loss = fr['non_member_losses'][idx]
+        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 = params['m_mean']
-        nm_mean = params['nm_mean']
-        m_std_v = params['m_std']
-        nm_std_v = params['nm_std']
+        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)
 
-    pred_label = "训练成员" if pred_member else "非训练成员"
-    pred_color = "🔴" if pred_member else "🟢"
-    actual_label = "训练成员" if is_member else "非训练成员"
-    actual_color = "🔴" if is_member else "🟢"
+    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:
-        verdict = "⚠️ **攻击成功: 发生了隐私泄露**"
-        verdict_detail = "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。"
+        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:
-        verdict = "✅ **判断正确**"
-        verdict_detail = "攻击者的判定与真实身份一致。"
+        v = "✅ **判断正确**"; vd = "攻击者的判定与真实身份一致。"
     else:
-        verdict = "❌ **攻击失误**"
-        verdict_detail = "攻击者的判定与真实身份不符。"
+        v = "❌ **攻击失误**"; vd = "攻击者的判定与真实身份不符。"
 
-    result_md = (
-        verdict + "\n\n" + verdict_detail + "\n\n"
+    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 | " + pred_color + " " + pred_label + " | " + actual_color + " " + actual_label + " |\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]
+    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���扰动 (s=0.02)": "baseline",
+}
+
+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,
+}
+
+
+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['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
+
+
 # ========================================
 # Interface
 # ========================================
 
 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;
-}
+.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;
-    transition: all 0.3s ease !important; background: transparent !important; border: none !important;
-}
-.tab-nav button:hover { color: #3b82f6 !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; }
@@ -397,219 +392,158 @@ body { background-color: #f0f4f8 !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 tr:last-child td { border-bottom: none !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; margin: 1.2em 0 !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; }
 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="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:
 
-    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及其对模型效用的影响。\n")
+    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")
 
     with gr.Tab("项目概览"):
         gr.Markdown(
-            "## ���究背景\n\n"
-            "大语言模型在教育领域广泛应用，训练过程中不可避免地接触到学生敏感数据。"
-            "**成员推理攻击 (MIA)** 能判断某条数据是否参与了模型训练，构成隐私威胁。\n\n"
-            "---\n\n"
+            "## 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"
-            "| 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，含姓名/学号/成绩等隐私字段 |\n"
-            "| 2. 基线模型训练 | Qwen2.5-Math-1.5B + LoRA | 标准微调，无任何防御措施 |\n"
-            "| 3. 标签平滑模型训练 | 两组不同平滑系数 | e=0.02(温和) 与 e=0.2(强力) 分别训练 |\n"
-            "| 4. MIA攻击测试 | 对全部模型及策略发起攻击 | 三个模型的Loss-based攻击 + 三组输出扰动测试 |\n"
-            "| 5. 效用评估 | 300道数学测试题 | 三个模型分别测试准确率 |\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"
+            "| 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"
-            "| 微调方法 | LoRA (r=8, alpha=16, target: q/k/v/o_proj) |\n"
-            "| 训练轮数 | 10 epochs |\n"
-            "| 数据总量 | " + data_size_str + " 条 (成员1000 + 非成员1000) |\n"
-            "| 训练模型数 | 3个 (基线 + 标签平滑x2) |\n"
-            "| 输出扰动测试 | 3组 (s=0.01/0.015/0.02，在基线模型上) |\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（训练集）与非成员数据1000条（对照组），每条均包含学生隐私字段。\n\n"
-                    "### 任务类型分布\n\n"
-                    "| 类型 | 数量 | 占比 | 说明 |\n"
-                    "|------|------|------|------|\n"
-                    "| 基础计算 | 800 | 40% | 加减乘除等基本运算 |\n"
-                    "| 应用题 | 600 | 30% | 实际场景的数学问题 |\n"
-                    "| 概念问答 | 400 | 20% | 数学概念理解 |\n"
-                    "| 错题订正 | 200 | 10% | 常见错误分析纠正 |\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")
         with gr.Row():
-            with gr.Column(scale=1):
-                gr.Markdown("**选择靶向数据池**")
-                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="")
+            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(scale=1):
-                gr.Markdown("**原始对话内容**")
+            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（模型或防御策略），系统将计算该样本的Loss值并判定成员身份。\n")
+        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.Row():
-            with gr.Column(scale=1):
-                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="模拟真实数据来源")
-                atk_idx = gr.Slider(0, 999, step=1, value=0, label="样本游标 ID (0-999)")
+            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(scale=1):
+            with gr.Column():
                 gr.Markdown("**攻击侦测控制台**")
-                atk_gauge = gr.Plot(label="Loss 分布雷达")
+                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 | 实验验证的优势 | 实验观察到的局限 |\n"
-            "|------|------|------|---------------|----------------|\n"
-            "| 标签平滑 | 训练期 | 软化训练标签，抑制过度记忆 | e=0.02时AUC降至" + f"{s002_auc:.4f}" + "，准确率" + f"{s002_acc:.1f}" + "% | 需重新训练；e过大时可能影响效用 |\n"
-            "| 输出扰动 | 推理期 | 对Loss添加高斯噪声 | s=0.02时AUC降至" + f"{op002_auc:.4f}" + "，准确率不变 | 仅遮蔽Loss统计信号，不改变模型记忆 |\n")
-        gr.Markdown("### AUC对比（全部策略）")
-        gr.Plot(value=make_auc_bar())
-        gr.Markdown("### Loss分布对比 - 三个模型（训练期防御效果）")
-        gr.Plot(value=make_loss_distribution())
-        gr.Markdown("### Loss分布对比 - 输出扰动（推理期防御效果）")
-        gr.Plot(value=make_perturb_loss_distribution())
-
-        tbl = (
-            "### 完整实验结果\n\n"
-            "| 策略 | 类型 | AUC | 准确率 | AUC变化 |\n"
-            "|------|------|-----|--------|--------|\n")
-        for k, name, cat in [('baseline', '基线 (无防御)', '--'), ('smooth_0.02', '标签平滑 (e=0.02)', '训练期'),
-                              ('smooth_0.2', '标签平滑 (e=0.2)', '训练期')]:
+        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
-                delta = "--" if k == 'baseline' else f"{a - bl_auc:+.4f}"
-                tbl += "| " + name + " | " + cat + " | " + f"{a:.4f}" + " | " + f"{acc:.1f}" + "% | " + delta + " |\n"
-        for k, name in [('perturbation_0.01', '输出扰动 (s=0.01)'), ('perturbation_0.015', '输���扰动 (s=0.015)'),
-                         ('perturbation_0.02', '输出扰动 (s=0.02)')]:
+                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']
-                delta = f"{a - bl_auc:+.4f}"
-                tbl += "| " + name + " | 推理期 | " + f"{a:.4f}" + " | " + f"{bl_acc:.1f}" + "% (不变) | " + delta + " |\n"
+                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"):
         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 (one-hot) 转换为软标签，降低模型对训练样本的过度拟合。\n\n"
+            "## 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。当 e=0 时退化为标准训练。\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，攻击风险较高 |\n"
-            "| e=0.02 | " + f"{s002_auc:.4f}" + " | " + f"{s002_acc:.1f}" + "% | 温和平滑，隐私与效用较好平衡 |\n"
-            "| e=0.2 | " + f"{s02_auc:.4f}" + " | " + f"{s02_acc:.1f}" + "% | 强力平滑，AUC显著下降 |\n\n"
-            "---\n\n"
-            "## 二、输出扰动 (Output Perturbation)\n\n"
-            "**类型**: 推理期防御\n\n"
-            "在推理阶段对模型返回的Loss值注入高斯噪声，使攻击者难以区分成员与非成员。\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"
-            "其中 s 为噪声标准差，控制扰动强度。\n\n"
-            "| 参数 | AUC | AUC降幅 | 准确率 |\n"
-            "|------|-----|---------|--------|\n"
-            "| 基线 (s=0) | " + 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}" + "% (不变) |\n"
-            "| s=0.015 | " + f"{op0015_auc:.4f}" + " | " + f"{bl_auc-op0015_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% (不变) |\n"
-            "| s=0.02 | " + f"{op002_auc:.4f}" + " | " + f"{bl_auc-op002_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% (不变) |\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"
-            "| 是否需要重训 | 是 | 否 |\n"
-            "| 对效用的影响 | 取决于平滑系数 | 无影响 |\n"
-            "| 防御原理 | 抑制过拟合，降低记忆 | 遮蔽Loss统计信号 |\n"
-            "| 部署难度 | 需训练阶段介入 | 推理阶段即插即用 |\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> 测试集: 300道数学题\n")
+        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")
         with gr.Row():
             with gr.Column():
-                gr.Markdown("### 准确率对比")
-                gr.Plot(value=make_accuracy_bar())
+                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(
-            "### 效用分析\n\n"
-            "| 策略 | 准确率 | AUC | 效用变化 | 分析 |\n"
-            "|------|--------|-----|---------|------|\n"
-            "| 基线 | " + f"{bl_acc:.1f}" + "% | " + f"{bl_auc:.4f}" + " | -- | 效用基准，隐私风险最高 |\n"
-            "| LS(e=0.02) | " + f"{s002_acc:.1f}" + "% | " + f"{s002_auc:.4f}" + " | " + f"{s002_acc-bl_acc:+.1f}" + "pp | 适度正则化提升泛化，准确率上升 |\n"
-            "| LS(e=0.2) | " + f"{s02_acc:.1f}" + "% | " + f"{s02_auc:.4f}" + " | " + f"{s02_acc-bl_acc:+.1f}" + "pp | 防御增强，效用仍可接受 |\n"
-            "| OP(s=0.01) | " + f"{bl_acc:.1f}" + "% | " + f"{op001_auc:.4f}" + " | 0 | 零效用损失 |\n"
-            "| OP(s=0.015) | " + f"{bl_acc:.1f}" + "% | " + f"{op0015_auc:.4f}" + " | 0 | 零效用损失 |\n"
-            "| OP(s=0.02) | " + f"{bl_acc:.1f}" + "% | " + f"{op002_auc:.4f}" + " | 0 | 零效用损失 |\n\n"
-            "> **关键发现**: 标签平滑 e=0.02 因正则化效应反而提升了泛化能力。"
-            "输出扰动在不影响效用的前提下实现了有效防御。两类策略在效用维度上呈现互补特性。\n")
+                gr.Markdown("### 隐私-效用权衡"); gr.Plot(value=make_tradeoff())
+        gr.Markdown("### 在线效用测试")
+        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分析"),
-                         ("fig3_defense_comparison_bar.png", "图3: 各防御策略MIA攻击AUC对比")]:
-            p = os.path.join(BASE_DIR, "figures", fn)
+        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)
             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=450); gr.Markdown("---")
 
     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成员推理攻击风险\n\n"
-            "基线模型AUC = **" + f"{bl_auc:.4f}" + "**，显著高于随机猜测 (0.5)。"
-            "成员平均Loss (" + f"{bl_m_mean:.4f}" + ") 低于非成员 (" + f"{bl_nm_mean:.4f}" + ")，"
-            "表明模型对训练数据产生了可被利用的记忆效应。\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"
-            "- e=0.02: AUC " + f"{bl_auc:.4f}" + " -> " + f"{s002_auc:.4f}" + "，准确率 " + f"{s002_acc:.1f}" + "%\n"
-            "- e=0.2: AUC " + f"{bl_auc:.4f}" + " -> " + f"{s02_auc:.4f}" + "，准确率 " + f"{s02_acc:.1f}" + "%\n\n"
-            "平滑系数需在保护强度与效用之间权衡，e=0.02取得较好平衡。\n\n---\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���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}" + "% (不变) |\n"
-            "| s=0.015 | " + f"{op0015_auc:.4f}" + " | " + f"{bl_auc-op0015_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% (不变) |\n"
-            "| s=0.02 | " + f"{op002_auc:.4f}" + " | " + f"{bl_auc-op002_auc:.4f}" + " | " + f"{bl_acc:.1f}" + "% (不变) |\n\n"
-            "零效用损失，适合已部署系统的后期隐私加固。\n\n---\n\n"
-            "### 四、隐私-效用权衡的定量分析\n\n"
-            "| 策略 | AUC | 准确率 | AUC变化 | 效用变化 |\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"
+            "### 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"
-            "两类策略在机制上互补：标签平滑从训练阶段降低记忆，输出扰动从推理阶段遮蔽统计信号。"
-            "实际部署中可根据场景灵活选择或组合。\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")