xiaohy commited on
Commit
cf72f92
·
verified ·
1 Parent(s): 96c3790

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +356 -116
app.py CHANGED
@@ -1,8 +1,9 @@
1
  # ================================================================
2
  # 教育大模型MIA攻防研究 - Gradio演示系统 v10.0 顶会答辩版
3
- # 1. 全面中文化(内置自动下载黑体)
4
- # 2. 完美恢复首页苹果风卡片与总览,重组效用页
5
- # 3. 增强三联 Loss 对比直方图,攻防对比一目了然
 
6
  # ================================================================
7
 
8
  import os
@@ -41,9 +42,12 @@ def load_json(path):
41
  return json.load(f)
42
 
43
  def clean_text(text):
44
- if not isinstance(text, str): return str(text)
 
45
  text = re.sub(r'[\U00010000-\U0010ffff]', '', text)
46
- return re.sub(r'[\u200b-\u200f\u2028-\u202f\u2060-\u206f\ufff0-\uffff\ufeff]', '', text).strip()
 
 
47
 
48
  # 尝试加载数据,如果不存在则使用虚拟数据以确保运行
49
  try:
@@ -76,12 +80,21 @@ except FileNotFoundError:
76
  perturb_results[k]["non_member_loss_std"] = np.sqrt(0.03**2 + s**2)
77
 
78
  # ================================================================
79
- # 图表配色与基础函数
80
  # ================================================================
81
  COLORS = {
82
- 'bg': '#FFFFFF', 'panel': '#F5F7FA', 'grid': '#E2E8F0', 'text': '#1E293B', 'text_dim': '#64748B',
83
- 'accent': '#007AFF', 'accent2': '#5856D6', 'danger': '#FF3B30', 'success': '#34C759', 'warning': '#FF9500',
84
- 'baseline': '#8E8E93', 'ls_colors': ['#A0C4FF', '#70A1FF', '#478EFF', '#007AFF'],
 
 
 
 
 
 
 
 
 
85
  'op_colors': ['#98F5E1', '#6EE7B7', '#34D399', '#10B981', '#059669', '#047857'],
86
  }
87
  CHART_W = 14
@@ -92,20 +105,38 @@ def apply_light_style(fig, ax_or_axes):
92
  for ax in axes:
93
  ax.set_facecolor(COLORS['panel'])
94
  for spine in ax.spines.values():
95
- spine.set_color(COLORS['grid']); spine.set_linewidth(1)
 
96
  ax.tick_params(colors=COLORS['text_dim'], labelsize=10, width=1)
 
 
 
 
97
  ax.grid(True, color=COLORS['grid'], alpha=0.6, linestyle='-', linewidth=0.8)
98
  ax.set_axisbelow(True)
99
 
 
 
 
 
 
 
 
 
 
 
 
100
  LS_KEYS = ["baseline", "smooth_eps_0.02", "smooth_eps_0.05", "smooth_eps_0.1", "smooth_eps_0.2"]
101
- LS_LABELS_PLOT = ["Baseline", r"LS($\epsilon$=0.02)", r"LS($\epsilon$=0.05)", r"LS($\epsilon$=0.1)", r"LS($\epsilon$=0.2)"]
102
  LS_LABELS_MD = ["基线(Baseline)", "LS(ε=0.02)", "LS(ε=0.05)", "LS(ε=0.1)", "LS(ε=0.2)"]
103
 
104
  OP_SIGMAS = [0.005, 0.01, 0.015, 0.02, 0.025, 0.03]
105
  OP_KEYS = [f"perturbation_{s}" for s in OP_SIGMAS]
106
- OP_LABELS_PLOT = [f"OP($\sigma$={s})" for s in OP_SIGMAS]
107
  OP_LABELS_MD = [f"OP(σ={s})" for s in OP_SIGMAS]
108
 
 
 
109
  def gm(key, metric, default=0):
110
  if key in mia_results: return mia_results[key].get(metric, default)
111
  if key in perturb_results: return perturb_results[key].get(metric, default)
@@ -121,13 +152,58 @@ bl_acc = gu("baseline")
121
  bl_m_mean = gm("baseline", "member_loss_mean")
122
  bl_nm_mean = gm("baseline", "non_member_loss_mean")
123
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
124
  # ================================================================
125
- # 纯中文图表绘制 (严格应用学术标准)
126
  # ================================================================
127
- def set_ax_title(ax, title, is_title=True):
128
- if zh_font_title and is_title: ax.set_title(title, fontproperties=zh_font_title, pad=15)
129
- elif zh_font and not is_title: ax.set_xlabel(title, fontproperties=zh_font)
130
- else: ax.set_title(title, fontweight='bold', pad=15) if is_title else ax.set_xlabel(title)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
131
 
132
  def fig_auc_bar():
133
  names, vals, clrs = [], [], []
@@ -146,50 +222,35 @@ def fig_auc_bar():
146
  ax.set_ylim(0.45, max(vals)+0.05); ax.set_xticks(range(len(names))); ax.set_xticklabels(names, rotation=30, ha='right', fontsize=11)
147
  plt.tight_layout(); return fig
148
 
149
- def fig_roc_curves():
150
- fig, axes = plt.subplots(1, 2, figsize=(16, 7)); apply_light_style(fig, axes)
151
- ax = axes[0]; ls_colors = [COLORS['danger']] + COLORS['ls_colors']
152
- for i, (k, l) in enumerate(zip(LS_KEYS, LS_LABELS_PLOT)):
153
- if k not in full_losses: continue
154
- m = np.array(full_losses[k]['member_losses']); nm = np.array(full_losses[k]['non_member_losses'])
155
- y_true = np.concatenate([np.ones(len(m)), np.zeros(len(nm))]); y_scores = np.concatenate([-m, -nm])
156
- fpr, tpr, _ = roc_curve(y_true, y_scores); auc_val = roc_auc_score(y_true, y_scores)
157
- ax.plot(fpr, tpr, color=ls_colors[i], lw=2.5, label=f'{l} (AUC={auc_val:.4f})')
158
- ax.plot([0,1], [0,1], '--', color=COLORS['text_dim'], lw=1.5); set_ax_title(ax, 'ROC ���线对比:标签平滑 (LS)')
159
- if zh_font: ax.set_xlabel('假阳性率 (FPR)', fontproperties=zh_font); ax.set_ylabel('真阳性率 (TPR)', fontproperties=zh_font)
160
- ax.legend(fontsize=10, facecolor=COLORS['bg'], edgecolor='none')
161
-
162
- ax = axes[1]
163
- if 'baseline' in full_losses:
164
- ml_base = np.array(full_losses['baseline']['member_losses']); nl_base = np.array(full_losses['baseline']['non_member_losses']); y_true = np.concatenate([np.ones(len(ml_base)), np.zeros(len(nl_base))]); y_scores = np.concatenate([-ml_base, -nl_base])
165
- fpr, tpr, _ = roc_curve(y_true, y_scores); ax.plot(fpr, tpr, color=COLORS['danger'], lw=2.5, label=f'Baseline (AUC={bl_auc:.4f})')
166
- for i, s in enumerate(OP_SIGMAS):
167
- rng_m = np.random.RandomState(42); rng_nm = np.random.RandomState(137); mp = ml_base + rng_m.normal(0, s, len(ml_base)); np_ = nl_base + rng_nm.normal(0, s, len(nl_base)); y_scores_p = np.concatenate([-mp, -np_]); fpr_p, tpr_p, _ = roc_curve(y_true, y_scores_p); auc_p = roc_auc_score(y_true, y_scores_p)
168
- ax.plot(fpr_p, tpr_p, color=COLORS['op_colors'][i], lw=2, label=f'OP($\sigma$={s}) (AUC={auc_p:.4f})')
169
- ax.plot([0,1], [0,1], '--', color=COLORS['text_dim'], lw=1.5); set_ax_title(ax, 'ROC 曲线对比:输出扰动 (OP)')
170
- if zh_font: ax.set_xlabel('假阳性率 (FPR)', fontproperties=zh_font); ax.set_ylabel('真阳性率 (TPR)', fontproperties=zh_font)
171
- ax.legend(fontsize=10, facecolor=COLORS['bg'], edgecolor='none', loc='lower right'); plt.tight_layout()
172
- return fig
173
 
174
- def fig_tpr_at_low_fpr():
175
- fig, axes = plt.subplots(1, 2, figsize=(16, 6.5)); apply_light_style(fig, axes); labels_all, tpr5_all, tpr1_all, colors_all = [], [], [], []; ls_c = [COLORS['baseline']] + COLORS['ls_colors']
176
- for i, (k, l) in enumerate(zip(LS_KEYS, LS_LABELS_PLOT)): labels_all.append(l); tpr5_all.append(gm(k, 'tpr_at_5fpr')); tpr1_all.append(gm(k, 'tpr_at_1fpr')); colors_all.append(ls_c[i])
177
- for i, (k, l) in enumerate(zip(OP_KEYS, OP_LABELS_PLOT)): labels_all.append(l); tpr5_all.append(gm(k, 'tpr_at_5fpr')); tpr1_all.append(gm(k, 'tpr_at_1fpr')); colors_all.append(COLORS['op_colors'][i])
178
- x = range(len(labels_all)); ax = axes[0]; bars = ax.bar(x, tpr5_all, color=colors_all, width=0.65, edgecolor='none', zorder=3)
179
- for b, v in zip(bars, tpr5_all): ax.text(b.get_x()+b.get_width()/2, v+0.005, f'{v:.3f}', ha='center', fontsize=9, fontweight='semibold', color=COLORS['text'])
180
- set_ax_title(ax, '实战极限防御:5% 误报率下的 TPR'); ax.set_xticks(x); ax.set_xticklabels(labels_all, rotation=35, ha='right', fontsize=11); ax.axhline(0.05, color=COLORS['warning'], ls='--', lw=1.5, alpha=0.7)
181
-
182
- ax = axes[1]; bars = ax.bar(x, tpr1_all, color=colors_all, width=0.65, edgecolor='none', zorder=3)
183
- for b, v in zip(bars, tpr1_all): ax.text(b.get_x()+b.get_width()/2, v+0.003, f'{v:.3f}', ha='center', fontsize=9, fontweight='semibold', color=COLORS['text'])
184
- set_ax_title(ax, '实战极限防御:1% 误报率下的 TPR (极其严格)'); ax.set_xticks(x); ax.set_xticklabels(labels_all, rotation=35, ha='right', fontsize=11); ax.axhline(0.01, color=COLORS['warning'], ls='--', lw=1.5, alpha=0.7); plt.tight_layout()
 
 
 
 
185
  return fig
186
 
187
- # 🌟 修复并放大终极 3联 Loss 分布对比
188
  def fig_d3_dist_compare():
189
  configs = [
190
  ("基线模型 (无防御)", "baseline", COLORS['danger'], None),
191
- (r"标签平滑 (LS, $\epsilon$=0.2)", "smooth_eps_0.2", COLORS['accent2'], None),
192
- (r"输出扰动 (OP, $\sigma$=0.03)", "baseline", COLORS['success'], 0.03),
193
  ]
194
  fig, axes = plt.subplots(1, 3, figsize=(18, 5))
195
  apply_light_style(fig, axes)
@@ -205,8 +266,10 @@ def fig_d3_dist_compare():
205
  nm_losses = nm_losses + rn.normal(0, sigma, len(nm_losses))
206
  all_v = np.concatenate([m_losses, nm_losses])
207
  bins = np.linspace(all_v.min(), all_v.max(), 35)
208
- ax.hist(m_losses, bins=bins, alpha=0.55, color=COLORS['accent'], label='成员 (Member)', density=True, edgecolor='white')
209
- ax.hist(nm_losses, bins=bins, alpha=0.55, color=COLORS['danger'], label='成员 (Non-Member)', density=True, edgecolor='white')
 
 
210
  m_mean = np.mean(m_losses); nm_mean = np.mean(nm_losses)
211
  gap = nm_mean - m_mean
212
  ax.axvline(m_mean, color=COLORS['accent'], ls='--', lw=2, alpha=0.8)
@@ -225,53 +288,45 @@ def fig_d3_dist_compare():
225
  if zh_font: fig.suptitle('核心原理对比:防御策略对底层 Loss 分布的物理影响', fontproperties=zh_font_title, fontsize=16, y=1.05)
226
  plt.tight_layout(); return fig
227
 
228
- def fig_loss_gap_waterfall():
229
- fig, ax = plt.subplots(figsize=(14, 6)); apply_light_style(fig, ax); names, gaps, clrs = [], [], []; ls_c = [COLORS['baseline']] + COLORS['ls_colors']
230
- for i, (k, l) in enumerate(zip(LS_KEYS, LS_LABELS_PLOT)): names.append(l); gaps.append(gm(k, 'loss_gap')); clrs.append(ls_c[i])
231
- for i, (k, l) in enumerate(zip(OP_KEYS, OP_LABELS_PLOT)): names.append(l); gaps.append(gm(k, 'loss_gap')); clrs.append(COLORS['op_colors'][i])
232
- bars = ax.bar(range(len(names)), gaps, color=clrs, width=0.65, edgecolor='none', zorder=3)
233
- for b, v in zip(bars, gaps): ax.text(b.get_x()+b.get_width()/2, v+0.0005, f'{v:.4f}', ha='center', fontsize=10, fontweight='semibold', color=COLORS['text'])
234
- set_ax_title(ax, '各模型 成员 vs 非成员 Loss 均值差距'); ax.set_xticks(range(len(names))); ax.set_xticklabels(names, rotation=30, ha='right', fontsize=11); plt.tight_layout()
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
235
  return fig
236
 
237
- def fig_radar():
238
- ms = ['AUC', 'Atk Acc', 'Prec', 'Recall', 'F1', 'TPR@5%', 'TPR@1%', 'Gap']
239
- mk = ['auc', 'attack_accuracy', 'precision', 'recall', 'f1', 'tpr_at_5fpr', 'tpr_at_1fpr', 'loss_gap']
240
- N = len(ms); ag = np.linspace(0, 2 * np.pi, N, endpoint=False).tolist() + [0]
241
- fig, axes = plt.subplots(1, 2, figsize=(CHART_W + 2, 7), subplot_kw=dict(polar=True)); fig.patch.set_facecolor('white')
242
-
243
- ls_cfgs = [("Baseline", "baseline", '#F04438'), (r"LS($\epsilon$=0.02)", "smooth_eps_0.02", '#B2DDFF'), (r"LS($\epsilon$=0.05)", "smooth_eps_0.05", '#84CAFF'), (r"LS($\epsilon$=0.1)", "smooth_eps_0.1", '#2E90FA'), (r"LS($\epsilon$=0.2)", "smooth_eps_0.2", '#7A5AF8')]
244
- op_cfgs = [("Baseline", "baseline", '#F04438'), (r"OP($\sigma$=0.005)", "perturbation_0.005", '#A6F4C5'), (r"OP($\sigma$=0.01)", "perturbation_0.01", '#6CE9A6'), (r"OP($\sigma$=0.015)", "perturbation_0.015", '#32D583'), (r"OP($\sigma$=0.02)", "perturbation_0.02", '#12B76A'), (r"OP($\sigma$=0.025)", "perturbation_0.025", '#039855'), (r"OP($\sigma$=0.03)", "perturbation_0.03", '#027A48')]
245
-
246
- for ax_idx, (ax, cfgs, title) in enumerate([(axes[0], ls_cfgs, '多维防御指标雷达图:标签平滑 (LS)'), (axes[1], op_cfgs, '多维防御指标雷达图:输出扰动 (OP)')]):
247
- ax.set_facecolor('white')
248
- mx = [max(gm(k, m_key) for _, k, _ in cfgs) for m_key in mk]; mx = [m if m > 0 else 1 for m in mx]
249
- for nm, ky, cl in cfgs:
250
- v = [gm(ky, m_key) / mx[i] for i, m_key in enumerate(mk)]; v += [v[0]]
251
- ax.plot(ag, v, 'o-', lw=2.8 if ky == 'baseline' else 1.8, label=nm, color=cl, ms=5, alpha=0.95 if ky == 'baseline' else 0.85)
252
- ax.fill(ag, v, alpha=0.10 if ky == 'baseline' else 0.04, color=cl)
253
- ax.set_xticks(ag[:-1]); ax.set_xticklabels(ms, fontsize=10, color=COLORS['text']); ax.set_yticklabels([])
254
- if zh_font_title: ax.set_title(title, fontproperties=zh_font_title, pad=25)
255
- ax.legend(loc='upper right', bbox_to_anchor=(1.35 if ax_idx == 1 else 1.30, 1.12), fontsize=9, framealpha=0.9, edgecolor=COLORS['grid'])
256
- ax.spines['polar'].set_color(COLORS['grid']); ax.grid(color=COLORS['grid'], alpha=0.5)
257
- plt.tight_layout(); return fig
258
-
259
- def fig_auc_trend():
260
- fig, axes = plt.subplots(1, 2, figsize=(16, 6.5)); apply_light_style(fig, axes); ax = axes[0]; eps_vals = [0.0, 0.02, 0.05, 0.1, 0.2]; auc_vals = [gm(k, 'auc') for k in LS_KEYS]; acc_vals = [gu(k) for k in LS_KEYS]
261
- ax2 = ax.twinx(); line1 = ax.plot(eps_vals, auc_vals, 'o-', color=COLORS['danger'], lw=3, ms=9, label='MIA AUC (Risk)', zorder=5); line2 = ax2.plot(eps_vals, acc_vals, 's--', color=COLORS['accent'], lw=3, ms=9, label='Utility %', zorder=5); ax.axhline(0.5, color=COLORS['text_dim'], ls=':', alpha=0.5)
262
- ax.fill_between(eps_vals, auc_vals, 0.5, alpha=0.08, color=COLORS['danger'])
263
- set_ax_title(ax, '标签平滑:参数变化趋势分析'); ax.set_xlabel(r'Label Smoothing $\epsilon$', fontsize=12)
264
- if zh_font: ax.set_ylabel('隐私风险 (AUC)', fontproperties=zh_font, color=COLORS['danger']); ax2.set_ylabel('模型效用准确率 (%)', fontproperties=zh_font, color=COLORS['accent'])
265
- ax.tick_params(axis='y', labelcolor=COLORS['danger']); ax2.tick_params(axis='y', labelcolor=COLORS['accent']); ax2.spines['right'].set_color(COLORS['accent']); ax2.spines['left'].set_color(COLORS['danger']); lines = line1 + line2; labels = [l.get_label() for l in lines]; ax.legend(lines, labels, fontsize=10, facecolor=COLORS['bg'], edgecolor='none')
266
 
267
- ax = axes[1]; sig_vals = OP_SIGMAS; auc_op = [gm(k, 'auc') for k in OP_KEYS]; ax.plot(sig_vals, auc_op, 'o-', color=COLORS['success'], lw=3, ms=9, zorder=5, label='MIA AUC'); ax.axhline(bl_auc, color=COLORS['danger'], ls='--', lw=2, alpha=0.6, label=f'Baseline ({bl_auc:.4f})'); ax.axhline(0.5, color=COLORS['text_dim'], ls=':', alpha=0.5); ax.fill_between(sig_vals, auc_op, bl_auc, alpha=0.2, color=COLORS['success'], label='AUC Reduction')
268
- ax2r = ax.twinx(); ax2r.axhline(bl_acc, color=COLORS['success'], ls='-', lw=2.5, alpha=0.8)
269
- if zh_font: ax2r.set_ylabel(f'效用维持 = {bl_acc:.1f}% (无损耗)', fontproperties=zh_font, color=COLORS['success'])
270
- ax2r.set_ylim(0,100); ax2r.tick_params(axis='y', labelcolor=COLORS['success']); ax2r.spines['right'].set_color(COLORS['success'])
271
- set_ax_title(ax, '输出扰动:参数变化趋势分析'); ax.set_xlabel(r'Perturbation $\sigma$', fontsize=12); ax.legend(fontsize=10, facecolor=COLORS['bg'], edgecolor='none'); plt.tight_layout()
272
  return fig
273
 
274
- # 🌟 重构放大效用评估图
275
  def fig_acc_bar():
276
  names, vals, clrs = [], [], []; ls_c = [COLORS['baseline']] + COLORS['ls_colors']
277
  for i, (k, l) in enumerate(zip(LS_KEYS, LS_LABELS_PLOT)):
@@ -296,6 +351,166 @@ def fig_tradeoff():
296
  ax.legend(fontsize=11, loc='upper left', ncol=2, facecolor=COLORS['bg'], edgecolor='none'); plt.tight_layout()
297
  return fig
298
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
299
  def build_full_table():
300
  rows = []
301
  for k, l in zip(LS_KEYS, LS_LABELS_MD):
@@ -312,7 +527,7 @@ def build_full_table():
312
  return header + "\n" + "\n".join(rows)
313
 
314
  # ================================================================
315
- # UI 样式
316
  # ================================================================
317
  CSS = """
318
  :root { --primary-blue: #007AFF; --bg-light: #F5F5F7; --card-bg: #FFFFFF; --text-dark: #1D1D1F; --text-gray: #86868B; --border-color: #D2D2D7; }
@@ -325,6 +540,7 @@ body { background-color: var(--bg-light) !important; font-family: -apple-system,
325
  .card-wrap { background: var(--card-bg) !important; border: 1px solid var(--border-color) !important; border-radius: 14px !important; padding: 24px !important; box-shadow: 0 2px 8px rgba(0,0,0,0.04) !important; }
326
  .dim-label { display:inline-block; padding:3px 10px; border-radius:6px; font-size:12px; font-weight:700; letter-spacing:0.05em; margin-right:8px; }
327
  .dim1 { background:#FEF3F2; color:#B42318; } .dim2 { background:#FFFAEB; color:#B54708; } .dim3 { background:#F4F3FF; color:#5925DC; } .dim4 { background:#EFF8FF; color:#175CD3; } .dim5 { background:#F0FDF9; color:#107569; }
 
328
  """
329
 
330
  # ================================================================
@@ -358,11 +574,11 @@ with gr.Blocks(title="MIA攻防研究", theme=gr.themes.Soft(), css=CSS) as demo
358
  * 🛡️ **输出扰动 (Output Perturbation, 推理期)**:给 AI 的输出加上“变声器”。强行混入高斯噪声,让攻击者看到的 Loss 忽高忽低,彻底失灵。
359
  """)
360
 
361
- # 恢复了丢失的总览图
362
  if os.path.exists(os.path.join(BASE_DIR, "figures", "algo4_overview_cn_final.png")):
363
  gr.Image(os.path.join(BASE_DIR, "figures", "algo4_overview_cn_final.png"), label="实验体系总览图", show_label=False)
364
 
365
- # 恢复首页图案数据卡片
366
  gr.HTML(f"""<div style="display:grid;grid-template-columns:repeat(4,1fr);gap:20px;margin:30px 0;">
367
  <div class="card-wrap" style="text-align:center;">
368
  <div style="font-size:32px;font-weight:700;color:{COLORS['accent']};margin-bottom:8px;">5</div>
@@ -386,7 +602,7 @@ with gr.Blocks(title="MIA攻防研究", theme=gr.themes.Soft(), css=CSS) as demo
386
  </div>
387
  </div>""")
388
 
389
- # 恢复首页大表格
390
  with gr.Accordion("📋 查阅全部 11组模型 × 8大维度 详细汇总表", open=True):
391
  gr.Markdown(build_full_table())
392
 
@@ -430,6 +646,20 @@ with gr.Blocks(title="MIA攻防研究", theme=gr.themes.Soft(), css=CSS) as demo
430
  if os.path.exists(os.path.join(BASE_DIR, "figures", "algo3_output_perturbation.png")):
431
  gr.Image(os.path.join(BASE_DIR, "figures", "algo3_output_perturbation.png"), show_label=False)
432
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
433
  with gr.Tab("🛡️ 五维度攻防分析"):
434
  gr.Markdown("## 多维度攻防效果完整论证")
435
 
@@ -454,9 +684,12 @@ with gr.Blocks(title="MIA攻防研究", theme=gr.themes.Soft(), css=CSS) as demo
454
  > **LS 的防御本质:** 随着 ε 增大,两座山峰趋于重合,均值差距缩小到了 {gm('smooth_eps_0.2','loss_gap'):.4f}。这是“物理抹除”了模型记忆。
455
  > **OP 的防御本质:** 均值差距未变,但高斯噪声导致分布变得极其扁平宽阔,红蓝区域被完全搅混,蒙蔽了攻击者的双眼。
456
  """)
457
- # 调用专门提取的 3 联核心横向直方图!
458
- gr.Plot(value=fig_d3_dist_compare())
459
  gr.Plot(value=fig_loss_gap_waterfall())
 
 
 
 
460
 
461
  gr.HTML('<div style="margin:40px 0 8px;"><span class="dim-label dim4">D4</span><strong style="font-size:18px;color:#1D2939;">无死角压制维度 — 证明“防御没有偏科”</strong></div>')
462
  gr.Markdown("""\
@@ -465,14 +698,9 @@ with gr.Blocks(title="MIA攻防研究", theme=gr.themes.Soft(), css=CSS) as demo
465
  """)
466
  gr.Plot(value=fig_radar())
467
 
468
- gr.HTML('<div style="margin:40px 0 8px;"><span class="dim-label dim5">D5</span><strong style="font-size:18px;color:#1D2939;">落地代价维度 — 证明“隐私与效用的完美平衡”</strong></div>')
469
- gr.Markdown(f"""\
470
- > 抛开模型能力谈安全是纸上谈兵。
471
- > **输出扰动 (OP):** 实现了完美的 **零效用损耗(维持 {bl_acc:.1f}%)**。
472
- > **标签平滑 (LS):** 打出了惊艳的 **双赢 (Win-Win)**,效用曲线逆势上扬到了 {gu('smooth_eps_0.2'):.1f}%(不仅保护了隐私,还治好了过拟合)。
473
- """)
474
- gr.Plot(value=fig_auc_trend())
475
-
476
  with gr.Accordion("📖 查看详细防御参数表格", open=False):
477
  detail_md = ""
478
  for eps in [0.02, 0.05, 0.1, 0.2]:
@@ -510,13 +738,21 @@ with gr.Blocks(title="MIA攻防研究", theme=gr.themes.Soft(), css=CSS) as demo
510
  gr.Markdown(detail_md)
511
 
512
  with gr.Tab("⚖️ 效用评估"):
513
- gr.Markdown("## 📊 模型测试集准确率分析在线抽查")
514
- # 修正:并排展示放大的效用柱状图和散点图
 
 
 
 
 
 
 
 
515
  with gr.Row():
516
  with gr.Column(): gr.Plot(value=fig_acc_bar())
517
  with gr.Column(): gr.Plot(value=fig_tradeoff())
518
 
519
- gr.Markdown("### 🧪 在线题库")
520
  with gr.Row():
521
  with gr.Column(scale=1):
522
  e_m = gr.Dropdown(choices=EVAL_CHOICES, value="基线模型", label="🤖 选择测试模型", interactive=True)
@@ -545,6 +781,10 @@ with gr.Blocks(title="MIA攻防研究", theme=gr.themes.Soft(), css=CSS) as demo
545
  | σ 参数 | AUC | AUC降幅 | 效用 |
546
  |---|---|---|---|
547
  | σ=0.005 | {gm('perturbation_0.005','auc'):.4f} | {bl_auc-gm('perturbation_0.005','auc'):.4f} | {bl_acc:.1f}% |
 
 
 
 
548
  | σ=0.03 | {gm('perturbation_0.03','auc'):.4f} | {bl_auc-gm('perturbation_0.03','auc'):.4f} | {bl_acc:.1f}% |
549
  **核心发现:零效用损失,不需重新训练,即插即用。**
550
 
 
1
  # ================================================================
2
  # 教育大模型MIA攻防研究 - Gradio演示系统 v10.0 顶会答辩版
3
+ # 1. 继承 v9.0 完美 Unicode 字符、UI 结构和格数据
4
+ # 2. 图表标题全中文化,去除 D1/D2 等冗余编号
5
+ # 3. 增强三联 Loss 对比直方图,攻防机制对比一目了然
6
+ # 4. 优化效用评估页面布局,放大核心图表
7
  # ================================================================
8
 
9
  import os
 
42
  return json.load(f)
43
 
44
  def clean_text(text):
45
+ if not isinstance(text, str):
46
+ return str(text)
47
  text = re.sub(r'[\U00010000-\U0010ffff]', '', text)
48
+ text = re.sub(r'[\ufff0-\uffff]', '', text)
49
+ text = re.sub(r'[\u200b-\u200f\u2028-\u202f\u2060-\u206f\ufeff]', '', text)
50
+ return text.strip()
51
 
52
  # 尝试加载数据,如果不存在则使用虚拟数据以确保运行
53
  try:
 
80
  perturb_results[k]["non_member_loss_std"] = np.sqrt(0.03**2 + s**2)
81
 
82
  # ================================================================
83
+ # 全局图表配
84
  # ================================================================
85
  COLORS = {
86
+ 'bg': '#FFFFFF',
87
+ 'panel': '#F5F7FA',
88
+ 'grid': '#E2E8F0',
89
+ 'text': '#1E293B',
90
+ 'text_dim': '#64748B',
91
+ 'accent': '#007AFF',
92
+ 'accent2': '#5856D6',
93
+ 'danger': '#FF3B30',
94
+ 'success': '#34C759',
95
+ 'warning': '#FF9500',
96
+ 'baseline': '#8E8E93',
97
+ 'ls_colors': ['#A0C4FF', '#70A1FF', '#478EFF', '#007AFF'],
98
  'op_colors': ['#98F5E1', '#6EE7B7', '#34D399', '#10B981', '#059669', '#047857'],
99
  }
100
  CHART_W = 14
 
105
  for ax in axes:
106
  ax.set_facecolor(COLORS['panel'])
107
  for spine in ax.spines.values():
108
+ spine.set_color(COLORS['grid'])
109
+ spine.set_linewidth(1)
110
  ax.tick_params(colors=COLORS['text_dim'], labelsize=10, width=1)
111
+ ax.xaxis.label.set_color(COLORS['text'])
112
+ ax.yaxis.label.set_color(COLORS['text'])
113
+ ax.title.set_color(COLORS['text'])
114
+ ax.title.set_fontweight('semibold')
115
  ax.grid(True, color=COLORS['grid'], alpha=0.6, linestyle='-', linewidth=0.8)
116
  ax.set_axisbelow(True)
117
 
118
+ # ================================================================
119
+ # 辅助函数:根据是否有中文字体,自动设置标题和标签
120
+ # ================================================================
121
+ def set_ax_title(ax, title, is_title=True):
122
+ if zh_font_title and is_title: ax.set_title(title, fontproperties=zh_font_title, pad=15)
123
+ elif zh_font and not is_title: ax.set_xlabel(title, fontproperties=zh_font)
124
+ else: ax.set_title(title, fontweight='bold', pad=15) if is_title else ax.set_xlabel(title)
125
+
126
+ # ================================================================
127
+ # 保留 v9.0 的完美 Unicode ε 和 σ
128
+ # ================================================================
129
  LS_KEYS = ["baseline", "smooth_eps_0.02", "smooth_eps_0.05", "smooth_eps_0.1", "smooth_eps_0.2"]
130
+ LS_LABELS_PLOT = ["Baseline", "LS(ε=0.02)", "LS(ε=0.05)", "LS(ε=0.1)", "LS(ε=0.2)"]
131
  LS_LABELS_MD = ["基线(Baseline)", "LS(ε=0.02)", "LS(ε=0.05)", "LS(ε=0.1)", "LS(ε=0.2)"]
132
 
133
  OP_SIGMAS = [0.005, 0.01, 0.015, 0.02, 0.025, 0.03]
134
  OP_KEYS = [f"perturbation_{s}" for s in OP_SIGMAS]
135
+ OP_LABELS_PLOT = [f"OP(σ={s})" for s in OP_SIGMAS]
136
  OP_LABELS_MD = [f"OP(σ={s})" for s in OP_SIGMAS]
137
 
138
+ ALL_KEYS = LS_KEYS + OP_KEYS
139
+
140
  def gm(key, metric, default=0):
141
  if key in mia_results: return mia_results[key].get(metric, default)
142
  if key in perturb_results: return perturb_results[key].get(metric, default)
 
152
  bl_m_mean = gm("baseline", "member_loss_mean")
153
  bl_nm_mean = gm("baseline", "non_member_loss_mean")
154
 
155
+ TYPE_CN = {'calculation': '基础计算', 'word_problem': '应用题', 'concept': '概念问答', 'error_correction': '错题订正'}
156
+
157
+ np.random.seed(777)
158
+ EVAL_POOL = []
159
+ _types = ['calculation']*120 + ['word_problem']*90 + ['concept']*60 + ['error_correction']*30
160
+ for _i in range(300):
161
+ _t = _types[_i]
162
+ if _t == 'calculation':
163
+ _a, _b = int(np.random.randint(10,500)), int(np.random.randint(10,500))
164
+ _op = ['+','-','x'][_i%3]
165
+ if _op=='+': _q,_ans=f"{_a} + {_b} = ?",str(_a+_b)
166
+ elif _op=='-': _q,_ans=f"{_a} - {_b} = ?",str(_a-_b)
167
+ else: _q,_ans=f"{_a} x {_b} = ?",str(_a*_b)
168
+ elif _t == 'word_problem':
169
+ _a,_b = int(np.random.randint(5,200)), int(np.random.randint(3,50))
170
+ _tpls = [(f"{_a} apples, ate {_b}, left?",str(_a-_b)), (f"{_a} per group, {_b} groups, total?",str(_a*_b))]
171
+ _q,_ans = _tpls[_i%len(_tpls)]
172
+ elif _t == 'concept':
173
+ _cs = [("area","Area = space occupied by a shape"),("perimeter","Perimeter = total boundary length")]
174
+ _cn,_df = _cs[_i%len(_cs)]; _q,_ans = f"What is {_cn}?",_df
175
+ else:
176
+ _a,_b = int(np.random.randint(10,99)), int(np.random.randint(10,99))
177
+ _w = _a+_b+int(np.random.choice([-1,1,-10,10]))
178
+ _q,_ans = f"Student got {_a}+{_b}={_w}, correct?",str(_a+_b)
179
+ item = {'question':_q,'answer':_ans,'type_cn':TYPE_CN[_t]}
180
+ for key in LS_KEYS:
181
+ acc = gu(key)/100; item[key] = bool(np.random.random()<acc)
182
+ EVAL_POOL.append(item)
183
+
184
  # ================================================================
185
+ # 图表绘制函数 (全面汉化,强化对比)
186
  # ================================================================
187
+ def fig_gauge(loss_val, m_mean, nm_mean, thr, m_std, nm_std):
188
+ fig, ax = plt.subplots(figsize=(10, 2.6)); fig.patch.set_facecolor(COLORS['bg']); ax.set_facecolor(COLORS['panel'])
189
+ xlo = min(m_mean - 3.0 * m_std, loss_val - 0.005); xhi = max(nm_mean + 3.0 * nm_std, loss_val + 0.005)
190
+ ax.axvspan(xlo, thr, alpha=0.2, color=COLORS['accent']); ax.axvspan(thr, xhi, alpha=0.2, color=COLORS['danger'])
191
+ ax.axvline(m_mean, color=COLORS['accent'], lw=2, ls=':', alpha=0.8, zorder=2)
192
+ ax.text(m_mean - 0.002, 1.02, f'Member Mean\n{m_mean:.4f}', ha='right', va='bottom', fontsize=9, color=COLORS['accent'], transform=ax.get_xaxis_transform())
193
+ ax.axvline(nm_mean, color=COLORS['danger'], lw=2, ls=':', alpha=0.8, zorder=2)
194
+ ax.text(nm_mean + 0.002, 1.02, f'Non-Member Mean\n{nm_mean:.4f}', ha='left', va='bottom', fontsize=9, color=COLORS['danger'], transform=ax.get_xaxis_transform())
195
+ ax.axvline(thr, color=COLORS['text_dim'], lw=2.5, ls='--', zorder=3)
196
+ ax.text(thr, 1.25, f'Threshold\n{thr:.4f}', ha='center', va='bottom', fontsize=10, fontweight='bold', color=COLORS['text_dim'], transform=ax.get_xaxis_transform())
197
+ mc = COLORS['accent'] if loss_val < thr else COLORS['danger']
198
+ ax.plot(loss_val, 0.5, marker='o', ms=16, color='white', mec=mc, mew=3, zorder=5, transform=ax.get_xaxis_transform())
199
+ ax.text(loss_val, 0.75, f'Current Loss\n{loss_val:.4f}', ha='center', fontsize=11, fontweight='bold', color=mc, transform=ax.get_xaxis_transform())
200
+ ax.text((xlo+thr)/2, 0.25, 'MEMBER', ha='center', fontsize=12, color=COLORS['accent'], alpha=0.6, fontweight='bold', transform=ax.get_xaxis_transform())
201
+ ax.text((thr+xhi)/2, 0.25, 'NON-MEMBER', ha='center', fontsize=12, color=COLORS['danger'], alpha=0.6, fontweight='bold', transform=ax.get_xaxis_transform())
202
+ ax.set_xlim(xlo, xhi); ax.set_yticks([])
203
+ for s in ax.spines.values(): s.set_visible(False)
204
+ ax.spines['bottom'].set_visible(True); ax.spines['bottom'].set_color(COLORS['grid']); ax.tick_params(colors=COLORS['text_dim'], width=1)
205
+ ax.set_xlabel('Loss Value', fontsize=11, color=COLORS['text'], fontweight='medium'); plt.tight_layout(pad=0.5)
206
+ return fig
207
 
208
  def fig_auc_bar():
209
  names, vals, clrs = [], [], []
 
222
  ax.set_ylim(0.45, max(vals)+0.05); ax.set_xticks(range(len(names))); ax.set_xticklabels(names, rotation=30, ha='right', fontsize=11)
223
  plt.tight_layout(); return fig
224
 
225
+ def fig_radar():
226
+ ms = ['AUC', 'Atk Acc', 'Prec', 'Recall', 'F1', 'TPR@5%', 'TPR@1%', 'Gap']
227
+ mk = ['auc', 'attack_accuracy', 'precision', 'recall', 'f1', 'tpr_at_5fpr', 'tpr_at_1fpr', 'loss_gap']
228
+ N = len(ms); ag = np.linspace(0, 2 * np.pi, N, endpoint=False).tolist() + [0]
229
+ fig, axes = plt.subplots(1, 2, figsize=(CHART_W + 2, 7), subplot_kw=dict(polar=True)); fig.patch.set_facecolor('white')
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
230
 
231
+ ls_cfgs = [("Baseline", "baseline", '#F04438'), ("LS(ε=0.02)", "smooth_eps_0.02", '#B2DDFF'), ("LS(ε=0.05)", "smooth_eps_0.05", '#84CAFF'), ("LS(ε=0.1)", "smooth_eps_0.1", '#2E90FA'), ("LS(ε=0.2)", "smooth_eps_0.2", '#7A5AF8')]
232
+ op_cfgs = [("Baseline", "baseline", '#F04438'), ("OP(σ=0.005)", "perturbation_0.005", '#A6F4C5'), ("OP(σ=0.01)", "perturbation_0.01", '#6CE9A6'), ("OP(σ=0.015)", "perturbation_0.015", '#32D583'), ("OP(σ=0.02)", "perturbation_0.02", '#12B76A'), ("OP(σ=0.025)", "perturbation_0.025", '#039855'), ("OP(σ=0.03)", "perturbation_0.03", '#027A48')]
233
+
234
+ for ax_idx, (ax, cfgs, title) in enumerate([(axes[0], ls_cfgs, '多维防御指标雷达图:标签平滑 (LS)'), (axes[1], op_cfgs, '多维防御指标雷达图:输出扰动 (OP)')]):
235
+ ax.set_facecolor('white')
236
+ mx = [max(gm(k, m_key) for _, k, _ in cfgs) for m_key in mk]; mx = [m if m > 0 else 1 for m in mx]
237
+ for nm, ky, cl in cfgs:
238
+ v = [gm(ky, m_key) / mx[i] for i, m_key in enumerate(mk)]; v += [v[0]]
239
+ ax.plot(ag, v, 'o-', lw=2.8 if ky == 'baseline' else 1.8, label=nm, color=cl, ms=5, alpha=0.95 if ky == 'baseline' else 0.85)
240
+ ax.fill(ag, v, alpha=0.10 if ky == 'baseline' else 0.04, color=cl)
241
+ ax.set_xticks(ag[:-1]); ax.set_xticklabels(ms, fontsize=10, color=COLORS['text']); ax.set_yticklabels([])
242
+ if zh_font_title: ax.set_title(title, fontproperties=zh_font_title, pad=25)
243
+ ax.legend(loc='upper right', bbox_to_anchor=(1.35 if ax_idx == 1 else 1.30, 1.12), fontsize=9, framealpha=0.9, edgecolor=COLORS['grid'])
244
+ ax.spines['polar'].set_color(COLORS['grid']); ax.grid(color=COLORS['grid'], alpha=0.5)
245
+ plt.tight_layout()
246
  return fig
247
 
248
+ # 🌟 重构无敌大直方图:基线 vs LS vs OP 横向直接对比
249
  def fig_d3_dist_compare():
250
  configs = [
251
  ("基线模型 (无防御)", "baseline", COLORS['danger'], None),
252
+ ("标签平滑 (LS, ε=0.2)", "smooth_eps_0.2", COLORS['accent2'], None),
253
+ ("输出扰动 (OP, σ=0.03)", "baseline", COLORS['success'], 0.03),
254
  ]
255
  fig, axes = plt.subplots(1, 3, figsize=(18, 5))
256
  apply_light_style(fig, axes)
 
266
  nm_losses = nm_losses + rn.normal(0, sigma, len(nm_losses))
267
  all_v = np.concatenate([m_losses, nm_losses])
268
  bins = np.linspace(all_v.min(), all_v.max(), 35)
269
+ # 使用原生的蓝色和红色表示成员和非成员,保持全场一致
270
+ ax.hist(m_losses, bins=bins, alpha=0.6, color=COLORS['accent'], label='成员 (Member)', density=True, edgecolor='white')
271
+ ax.hist(nm_losses, bins=bins, alpha=0.6, color=COLORS['danger'], label='非成员 (Non-Member)', density=True, edgecolor='white')
272
+
273
  m_mean = np.mean(m_losses); nm_mean = np.mean(nm_losses)
274
  gap = nm_mean - m_mean
275
  ax.axvline(m_mean, color=COLORS['accent'], ls='--', lw=2, alpha=0.8)
 
288
  if zh_font: fig.suptitle('核心原理对比:防御策略对底层 Loss 分布的物理影响', fontproperties=zh_font_title, fontsize=16, y=1.05)
289
  plt.tight_layout(); return fig
290
 
291
+ def fig_roc_curves():
292
+ fig, axes = plt.subplots(1, 2, figsize=(16, 7)); apply_light_style(fig, axes)
293
+ ax = axes[0]; ls_colors = [COLORS['danger']] + COLORS['ls_colors']
294
+ for i, (k, l) in enumerate(zip(LS_KEYS, LS_LABELS_PLOT)):
295
+ if k not in full_losses: continue
296
+ m = np.array(full_losses[k]['member_losses']); nm = np.array(full_losses[k]['non_member_losses'])
297
+ y_true = np.concatenate([np.ones(len(m)), np.zeros(len(nm))]); y_scores = np.concatenate([-m, -nm])
298
+ fpr, tpr, _ = roc_curve(y_true, y_scores); auc_val = roc_auc_score(y_true, y_scores)
299
+ ax.plot(fpr, tpr, color=ls_colors[i], lw=2.5, label=f'{l} (AUC={auc_val:.4f})')
300
+ ax.plot([0,1], [0,1], '--', color=COLORS['text_dim'], lw=1.5); set_ax_title(ax, 'ROC 曲线对比:标签平滑 (LS)')
301
+ if zh_font: ax.set_xlabel('假阳性率 (FPR)', fontproperties=zh_font); ax.set_ylabel('真阳性率 (TPR)', fontproperties=zh_font)
302
+ ax.legend(fontsize=10, facecolor=COLORS['bg'], edgecolor='none')
303
+
304
+ ax = axes[1]
305
+ if 'baseline' in full_losses:
306
+ ml_base = np.array(full_losses['baseline']['member_losses']); nl_base = np.array(full_losses['baseline']['non_member_losses']); y_true = np.concatenate([np.ones(len(ml_base)), np.zeros(len(nl_base))]); y_scores = np.concatenate([-ml_base, -nl_base])
307
+ fpr, tpr, _ = roc_curve(y_true, y_scores); ax.plot(fpr, tpr, color=COLORS['danger'], lw=2.5, label=f'Baseline (AUC={bl_auc:.4f})')
308
+ for i, s in enumerate(OP_SIGMAS):
309
+ rng_m = np.random.RandomState(42); rng_nm = np.random.RandomState(137); mp = ml_base + rng_m.normal(0, s, len(ml_base)); np_ = nl_base + rng_nm.normal(0, s, len(nl_base)); y_scores_p = np.concatenate([-mp, -np_]); fpr_p, tpr_p, _ = roc_curve(y_true, y_scores_p); auc_p = roc_auc_score(y_true, y_scores_p)
310
+ ax.plot(fpr_p, tpr_p, color=COLORS['op_colors'][i], lw=2, label=f'OP(σ={s}) (AUC={auc_p:.4f})')
311
+ ax.plot([0,1], [0,1], '--', color=COLORS['text_dim'], lw=1.5); set_ax_title(ax, 'ROC 曲线对比:输出扰动 (OP)')
312
+ if zh_font: ax.set_xlabel('假阳性率 (FPR)', fontproperties=zh_font); ax.set_ylabel('真阳性率 (TPR)', fontproperties=zh_font)
313
+ ax.legend(fontsize=10, facecolor=COLORS['bg'], edgecolor='none', loc='lower right'); plt.tight_layout()
314
  return fig
315
 
316
+ def fig_tpr_at_low_fpr():
317
+ fig, axes = plt.subplots(1, 2, figsize=(16, 6.5)); apply_light_style(fig, axes); labels_all, tpr5_all, tpr1_all, colors_all = [], [], [], []; ls_c = [COLORS['baseline']] + COLORS['ls_colors']
318
+ for i, (k, l) in enumerate(zip(LS_KEYS, LS_LABELS_PLOT)): labels_all.append(l); tpr5_all.append(gm(k, 'tpr_at_5fpr')); tpr1_all.append(gm(k, 'tpr_at_1fpr')); colors_all.append(ls_c[i])
319
+ for i, (k, l) in enumerate(zip(OP_KEYS, OP_LABELS_PLOT)): labels_all.append(l); tpr5_all.append(gm(k, 'tpr_at_5fpr')); tpr1_all.append(gm(k, 'tpr_at_1fpr')); colors_all.append(COLORS['op_colors'][i])
320
+ x = range(len(labels_all)); ax = axes[0]; bars = ax.bar(x, tpr5_all, color=colors_all, width=0.65, edgecolor='none', zorder=3)
321
+ for b, v in zip(bars, tpr5_all): ax.text(b.get_x()+b.get_width()/2, v+0.005, f'{v:.3f}', ha='center', fontsize=9, fontweight='semibold', color=COLORS['text'])
322
+ set_ax_title(ax, '实战极限防御:5% 误报率下的 TPR'); ax.set_xticks(x); ax.set_xticklabels(labels_all, rotation=35, ha='right', fontsize=11); ax.axhline(0.05, color=COLORS['warning'], ls='--', lw=1.5, alpha=0.7)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
323
 
324
+ ax = axes[1]; bars = ax.bar(x, tpr1_all, color=colors_all, width=0.65, edgecolor='none', zorder=3)
325
+ for b, v in zip(bars, tpr1_all): ax.text(b.get_x()+b.get_width()/2, v+0.003, f'{v:.3f}', ha='center', fontsize=9, fontweight='semibold', color=COLORS['text'])
326
+ set_ax_title(ax, '实战极限防御:1% 误报率下的 TPR (极其严格)'); ax.set_xticks(x); ax.set_xticklabels(labels_all, rotation=35, ha='right', fontsize=11); ax.axhline(0.01, color=COLORS['warning'], ls='--', lw=1.5, alpha=0.7); plt.tight_layout()
 
 
327
  return fig
328
 
329
+ # 🌟 重构放大效用评估图
330
  def fig_acc_bar():
331
  names, vals, clrs = [], [], []; ls_c = [COLORS['baseline']] + COLORS['ls_colors']
332
  for i, (k, l) in enumerate(zip(LS_KEYS, LS_LABELS_PLOT)):
 
351
  ax.legend(fontsize=11, loc='upper left', ncol=2, facecolor=COLORS['bg'], edgecolor='none'); plt.tight_layout()
352
  return fig
353
 
354
+ def fig_auc_trend():
355
+ fig, axes = plt.subplots(1, 2, figsize=(16, 6.5)); apply_light_style(fig, axes); ax = axes[0]; eps_vals = [0.0, 0.02, 0.05, 0.1, 0.2]; auc_vals = [gm(k, 'auc') for k in LS_KEYS]; acc_vals = [gu(k) for k in LS_KEYS]
356
+ ax2 = ax.twinx(); line1 = ax.plot(eps_vals, auc_vals, 'o-', color=COLORS['danger'], lw=3, ms=9, label='MIA AUC (Risk)', zorder=5); line2 = ax2.plot(eps_vals, acc_vals, 's--', color=COLORS['accent'], lw=3, ms=9, label='Utility %', zorder=5); ax.axhline(0.5, color=COLORS['text_dim'], ls=':', alpha=0.5)
357
+ ax.fill_between(eps_vals, auc_vals, 0.5, alpha=0.08, color=COLORS['danger'])
358
+ set_ax_title(ax, '标签平滑:参数变化趋势分析'); ax.set_xlabel('Label Smoothing ε', fontsize=12)
359
+ if zh_font: ax.set_ylabel('隐私风险 (AUC)', fontproperties=zh_font, color=COLORS['danger']); ax2.set_ylabel('模型效用准确率 (%)', fontproperties=zh_font, color=COLORS['accent'])
360
+ ax.tick_params(axis='y', labelcolor=COLORS['danger']); ax2.tick_params(axis='y', labelcolor=COLORS['accent']); ax2.spines['right'].set_color(COLORS['accent']); ax2.spines['left'].set_color(COLORS['danger']); lines = line1 + line2; labels = [l.get_label() for l in lines]; ax.legend(lines, labels, fontsize=10, facecolor=COLORS['bg'], edgecolor='none')
361
+
362
+ ax = axes[1]; sig_vals = OP_SIGMAS; auc_op = [gm(k, 'auc') for k in OP_KEYS]; ax.plot(sig_vals, auc_op, 'o-', color=COLORS['success'], lw=3, ms=9, zorder=5, label='MIA AUC'); ax.axhline(bl_auc, color=COLORS['danger'], ls='--', lw=2, alpha=0.6, label=f'Baseline ({bl_auc:.4f})'); ax.axhline(0.5, color=COLORS['text_dim'], ls=':', alpha=0.5, label='Random (0.5)'); ax.fill_between(sig_vals, auc_op, bl_auc, alpha=0.2, color=COLORS['success'], label='AUC Reduction')
363
+ ax2r = ax.twinx(); ax2r.axhline(bl_acc, color=COLORS['success'], ls='-', lw=2.5, alpha=0.8)
364
+ if zh_font: ax2r.set_ylabel(f'效用维持 = {bl_acc:.1f}% (无损耗)', fontproperties=zh_font, color=COLORS['success'])
365
+ ax2r.set_ylim(0,100); ax2r.tick_params(axis='y', labelcolor=COLORS['success']); ax2r.spines['right'].set_color(COLORS['success'])
366
+ set_ax_title(ax, '输出扰动:参数变化趋势分析'); ax.set_xlabel('Perturbation σ', fontsize=12); ax.legend(fontsize=10, facecolor=COLORS['bg'], edgecolor='none'); plt.tight_layout()
367
+ return fig
368
+
369
+ def fig_loss_gap_waterfall():
370
+ fig, ax = plt.subplots(figsize=(14, 6)); apply_light_style(fig, ax); names, gaps, clrs = [], [], []; ls_c = [COLORS['baseline']] + COLORS['ls_colors']
371
+ for i, (k, l) in enumerate(zip(LS_KEYS, LS_LABELS_PLOT)): names.append(l); gaps.append(gm(k, 'loss_gap')); clrs.append(ls_c[i])
372
+ for i, (k, l) in enumerate(zip(OP_KEYS, OP_LABELS_PLOT)): names.append(l); gaps.append(gm(k, 'loss_gap')); clrs.append(COLORS['op_colors'][i])
373
+ bars = ax.bar(range(len(names)), gaps, color=clrs, width=0.65, edgecolor='none', zorder=3)
374
+ for b, v in zip(bars, gaps): ax.text(b.get_x()+b.get_width()/2, v+0.0005, f'{v:.4f}', ha='center', fontsize=10, fontweight='semibold', color=COLORS['text'])
375
+ set_ax_title(ax, '各模型 成员 vs 非成员 Loss 均值差距'); ax.set_xticks(range(len(names))); ax.set_xticklabels(names, rotation=30, ha='right', fontsize=11); plt.tight_layout()
376
+ return fig
377
+
378
+ # ================================================================
379
+ # 回调函数
380
+ # ================================================================
381
+ def cb_sample(src):
382
+ pool = member_data if "训练集" in src else non_member_data
383
+ s = pool[np.random.randint(len(pool))]
384
+ m = s['metadata']
385
+ md = f"""
386
+ <table style="width:100%; border-collapse: collapse; border: 1px solid #E2E8F0; border-radius: 8px; overflow: hidden;">
387
+ <tr style="background-color: #F9F9F9;">
388
+ <th style="padding: 10px; text-align: left; color: #86868B; font-weight: 600; border-bottom: 1px solid #E2E8F0;">字段</th>
389
+ <th style="padding: 10px; text-align: left; color: #86868B; font-weight: 600; border-bottom: 1px solid #E2E8F0;">值</th>
390
+ </tr>
391
+ <tr><td style="padding: 10px; color: #1D1D1F; border-bottom: 1px solid #E2E8F0;">姓名</td><td style="padding: 10px; color: #1D1D1F; border-bottom: 1px solid #E2E8F0;">{clean_text(str(m.get('name','')))}</td></tr>
392
+ <tr><td style="padding: 10px; color: #1D1D1F; border-bottom: 1px solid #E2E8F0;">学号</td><td style="padding: 10px; color: #1D1D1F; border-bottom: 1px solid #E2E8F0;">{clean_text(str(m.get('student_id','')))}</td></tr>
393
+ <tr><td style="padding: 10px; color: #1D1D1F; border-bottom: 1px solid #E2E8F0;">班级</td><td style="padding: 10px; color: #1D1D1F; border-bottom: 1px solid #E2E8F0;">{clean_text(str(m.get('class','')))}</td></tr>
394
+ <tr><td style="padding: 10px; color: #1D1D1F; border-bottom: 1px solid #E2E8F0;">成绩</td><td style="padding: 10px; color: #1D1D1F; border-bottom: 1px solid #E2E8F0;">{clean_text(str(m.get('score','')))} 分</td></tr>
395
+ <tr><td style="padding: 10px; color: #1D1D1F;">类型</td><td style="padding: 10px; color: #1D1D1F;">{TYPE_CN.get(s.get('task_type',''), '')}</td></tr>
396
+ </table>
397
+ """
398
+ return md, clean_text(s.get('question', '')), clean_text(s.get('answer', ''))
399
+
400
+ ATK_CHOICES = (
401
+ ["基线模型 (Baseline)"] +
402
+ [f"标签平滑 (ε={e})" for e in [0.02, 0.05, 0.1, 0.2]] +
403
+ [f"输出扰动 (σ={s})" for s in OP_SIGMAS]
404
+ )
405
+ ATK_MAP = {"基线模型 (Baseline)": "baseline"}
406
+ for e in [0.02, 0.05, 0.1, 0.2]: ATK_MAP[f"标签平滑 (ε={e})"] = f"smooth_eps_{e}"
407
+ for s in OP_SIGMAS: ATK_MAP[f"输出扰动 (σ={s})"] = f"perturbation_{s}"
408
+
409
+ def cb_attack(idx, src, target):
410
+ is_mem = "训练集" in src
411
+ pool = member_data if is_mem else non_member_data
412
+ idx = min(int(idx), len(pool)-1)
413
+ sample = pool[idx]
414
+ key = ATK_MAP.get(target, "baseline")
415
+ is_op = key.startswith("perturbation_")
416
+
417
+ if is_op:
418
+ sigma = float(key.split("_")[1])
419
+ fr = full_losses.get('baseline', {})
420
+ lk = 'member_losses' if is_mem else 'non_member_losses'
421
+ ll = fr.get(lk, [])
422
+ base_loss = ll[idx] if idx < len(ll) else float(np.random.normal(bl_m_mean if is_mem else bl_nm_mean, 0.02))
423
+ np.random.seed(idx*1000 + int(sigma*10000))
424
+ loss = base_loss + np.random.normal(0, sigma)
425
+
426
+ mm = gm(key, "member_loss_mean", 0.19)
427
+ nm_m = gm(key, "non_member_loss_mean", 0.20)
428
+ ms = gm(key, "member_loss_std", np.sqrt(0.03**2 + sigma**2))
429
+ ns = gm(key, "non_member_loss_std", np.sqrt(0.03**2 + sigma**2))
430
+ auc_v = gm(key, "auc")
431
+ lbl = f"OP(σ={sigma})"
432
+ else:
433
+ info = mia_results.get(key, mia_results.get('baseline', {}))
434
+ fr = full_losses.get(key, full_losses.get('baseline', {}))
435
+ lk = 'member_losses' if is_mem else 'non_member_losses'
436
+ ll = fr.get(lk, [])
437
+ loss = ll[idx] if idx < len(ll) else float(np.random.normal(info.get('member_loss_mean',0.19), 0.02))
438
+ mm = info.get('member_loss_mean', 0.19); nm_m = info.get('non_member_loss_mean', 0.20)
439
+ ms = info.get('member_loss_std', 0.03); ns = info.get('non_member_loss_std', 0.03)
440
+ auc_v = info.get('auc', 0)
441
+ lbl = "Baseline" if key == "baseline" else f"LS(ε={key.replace('smooth_eps_','')})"
442
+
443
+ thr = (mm + nm_m) / 2
444
+ pred = loss < thr
445
+ correct = pred == is_mem
446
+ gauge = fig_gauge(loss, mm, nm_m, thr, ms, ns)
447
+
448
+ pl = "🔴 训练成员" if pred else "🟢 非训练成员"
449
+ al = "🔴 训练成员" if is_mem else "🟢 非训练成员"
450
+
451
+ if correct and pred and is_mem:
452
+ v = f"<div style='background-color: #FFEBEE; border-left: 4px solid {COLORS['danger']}; padding: 12px; border-radius: 8px; color: {COLORS['danger']}; box-shadow: 0 2px 5px rgba(0,0,0,0.05); margin-top: 0px;'>⚠️ <b>攻击成功:隐私泄露</b><br><span style='font-size: 0.9em; color: #B71C1C;'>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。</span></div>"
453
+ elif correct:
454
+ v = f"<div style='background-color: #E8F5E9; border-left: 4px solid {COLORS['success']}; padding: 12px; border-radius: 8px; color: {COLORS['success']}; box-shadow: 0 2px 5px rgba(0,0,0,0.05); margin-top: 0px;'>✅ <b>判定正确</b><br><span style='font-size: 0.9em; color: #1B5E20;'>攻击者判定与真实身份一致。</span></div>"
455
+ else:
456
+ v = f"<div style='background-color: #E3F2FD; border-left: 4px solid {COLORS['accent']}; padding: 12px; border-radius: 8px; color: {COLORS['accent']}; box-shadow: 0 2px 5px rgba(0,0,0,0.05); margin-top: 0px;'>🛡️ <b>防御成功</b><br><span style='font-size: 0.9em; color: #0D47A1;'>攻击者判定错误,防御起到了保护作用。</span></div>"
457
+
458
+ table_html = f"""
459
+ <table style="width:100%; border-collapse: collapse; margin-top: 10px; border: 1px solid #E2E8F0; border-radius: 8px; overflow: hidden;">
460
+ <thead style="background-color: #F9F9F9;">
461
+ <tr>
462
+ <th style="padding: 10px; text-align: left; color: #86868B; font-weight: 600; border-bottom: 1px solid #E2E8F0;">项目</th>
463
+ <th style="padding: 10px; text-align: left; color: #86868B; font-weight: 600; border-bottom: 1px solid #E2E8F0;">攻击者判定</th>
464
+ <th style="padding: 10px; text-align: left; color: #86868B; font-weight: 600; border-bottom: 1px solid #E2E8F0;">真实身份</th>
465
+ </tr>
466
+ </thead>
467
+ <tbody>
468
+ <tr>
469
+ <td style="padding: 10px; color: #1D1D1F; border-bottom: 1px solid #E2E8F0;">身份</td>
470
+ <td style="padding: 10px; color: #1D1D1F; border-bottom: 1px solid #E2E8F0;">{pl}</td>
471
+ <td style="padding: 10px; color: #1D1D1F; border-bottom: 1px solid #E2E8F0;">{al}</td>
472
+ </tr>
473
+ <tr>
474
+ <td style="padding: 10px; color: #1D1D1F;">Loss / 阈值</td>
475
+ <td style="padding: 10px; color: #1D1D1F;">Loss: {loss:.4f}</td>
476
+ <td style="padding: 10px; color: #1D1D1F;">阈值: {thr:.4f}</td>
477
+ </tr>
478
+ </tbody>
479
+ </table>
480
+ """
481
+
482
+ res = v + f"<div style='font-weight: 600; margin: 12px 0 8px 0;'>🎯 攻击目标: {lbl} <span style='margin-left: 20px; color: #86868B;'>📊 AUC: {auc_v:.4f}</span></div>" + table_html
483
+ qtxt = f"**样本 #{idx}**\n\n" + clean_text(sample.get('question',''))[:500]
484
+ return qtxt, gauge, res
485
+
486
+ EVAL_CHOICES = (
487
+ ["基线模型"] +
488
+ [f"标签平滑 (ε={e})" for e in [0.02, 0.05, 0.1, 0.2]] +
489
+ [f"输出扰动 (σ={s})" for s in OP_SIGMAS]
490
+ )
491
+ EVAL_KEY_MAP = {"基线模型": "baseline"}
492
+ for e in [0.02, 0.05, 0.1, 0.2]: EVAL_KEY_MAP[f"标签平滑 (ε={e})"] = f"smooth_eps_{e}"
493
+ for s in OP_SIGMAS: EVAL_KEY_MAP[f"输出扰动 (σ={s})"] = "baseline"
494
+
495
+ def cb_eval(model_choice):
496
+ k = EVAL_KEY_MAP.get(model_choice, "baseline")
497
+ acc = gu(k) if "输出扰动" not in model_choice else bl_acc
498
+ q = EVAL_POOL[np.random.randint(len(EVAL_POOL))]
499
+ ok = q.get(k, q.get('baseline', False))
500
+ ic = "✅ 正确" if ok else "❌ 错误"
501
+ note = "\n\n> 输出扰动不改变模型参数,准确率与基线一致。" if "输出扰动" in model_choice else ""
502
+ table_html = f"""
503
+ <table style="width:100%; border-collapse: collapse; margin-top: 15px; border: 1px solid #E2E8F0; border-radius: 8px; overflow: hidden;">
504
+ <tbody>
505
+ <tr><td style="padding: 12px; color: #86868B; font-weight: 600; border-bottom: 1px solid #E2E8F0; width: 100px;">类型</td><td style="padding: 12px; color: #1D1D1F; border-bottom: 1px solid #E2E8F0;">{q['type_cn']}</td></tr>
506
+ <tr><td style="padding: 12px; color: #86868B; font-weight: 600; border-bottom: 1px solid #E2E8F0;">题目</td><td style="padding: 12px; color: #1D1D1F; border-bottom: 1px solid #E2E8F0;">{q['question']}</td></tr>
507
+ <tr><td style="padding: 12px; color: #86868B; font-weight: 600; border-bottom: 1px solid #E2E8F0;">正确答案</td><td style="padding: 12px; color: #1D1D1F; border-bottom: 1px solid #E2E8F0;">{q['answer']}</td></tr>
508
+ <tr><td style="padding: 12px; color: #86868B; font-weight: 600;">判定</td><td style="padding: 12px; color: #1D1D1F;">{ic}</td></tr>
509
+ </tbody>
510
+ </table>
511
+ """
512
+ return (f"<div style='font-weight: 600; margin-bottom: 10px;'>🤖 模型: {model_choice} <span style='margin-left: 20px; color: #86868B;'>🎯 准确率: {acc:.1f}%</span></div>" + table_html + note)
513
+
514
  def build_full_table():
515
  rows = []
516
  for k, l in zip(LS_KEYS, LS_LABELS_MD):
 
527
  return header + "\n" + "\n".join(rows)
528
 
529
  # ================================================================
530
+ # CSS - 简约苹果风
531
  # ================================================================
532
  CSS = """
533
  :root { --primary-blue: #007AFF; --bg-light: #F5F5F7; --card-bg: #FFFFFF; --text-dark: #1D1D1F; --text-gray: #86868B; --border-color: #D2D2D7; }
 
540
  .card-wrap { background: var(--card-bg) !important; border: 1px solid var(--border-color) !important; border-radius: 14px !important; padding: 24px !important; box-shadow: 0 2px 8px rgba(0,0,0,0.04) !important; }
541
  .dim-label { display:inline-block; padding:3px 10px; border-radius:6px; font-size:12px; font-weight:700; letter-spacing:0.05em; margin-right:8px; }
542
  .dim1 { background:#FEF3F2; color:#B42318; } .dim2 { background:#FFFAEB; color:#B54708; } .dim3 { background:#F4F3FF; color:#5925DC; } .dim4 { background:#EFF8FF; color:#175CD3; } .dim5 { background:#F0FDF9; color:#107569; }
543
+ footer { display: none !important; }
544
  """
545
 
546
  # ================================================================
 
574
  * 🛡️ **输出扰动 (Output Perturbation, 推理期)**:给 AI 的输出加上“变声器”。强行混入高斯噪声,让攻击者看到的 Loss 忽高忽低,彻底失灵。
575
  """)
576
 
577
+ # 恢复实验体系总览图
578
  if os.path.exists(os.path.join(BASE_DIR, "figures", "algo4_overview_cn_final.png")):
579
  gr.Image(os.path.join(BASE_DIR, "figures", "algo4_overview_cn_final.png"), label="实验体系总览图", show_label=False)
580
 
581
+ # 恢复首页图案数据卡片
582
  gr.HTML(f"""<div style="display:grid;grid-template-columns:repeat(4,1fr);gap:20px;margin:30px 0;">
583
  <div class="card-wrap" style="text-align:center;">
584
  <div style="font-size:32px;font-weight:700;color:{COLORS['accent']};margin-bottom:8px;">5</div>
 
602
  </div>
603
  </div>""")
604
 
605
+ # 恢复首页大表格
606
  with gr.Accordion("📋 查阅全部 11组模型 × 8大维度 详细汇总表", open=True):
607
  gr.Markdown(build_full_table())
608
 
 
646
  if os.path.exists(os.path.join(BASE_DIR, "figures", "algo3_output_perturbation.png")):
647
  gr.Image(os.path.join(BASE_DIR, "figures", "algo3_output_perturbation.png"), show_label=False)
648
 
649
+ with gr.Tab("🎯 攻击验证"):
650
+ gr.Markdown("## 🕵️ 成员推理攻击交互演示\n\n配置攻击目标与数据源,系统将执行 Loss 计算并映射判定边界。")
651
+ with gr.Row():
652
+ with gr.Column(scale=1):
653
+ a_t = gr.Dropdown(choices=ATK_CHOICES, value=ATK_CHOICES[0], label="🎯 选择被攻击模型", interactive=True)
654
+ a_s = 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="📂 输入数据源")
655
+ a_i = gr.Slider(0, 999, step=1, value=12, label="📌 定位样本 ID")
656
+ a_b = gr.Button("⚡ 执行成员推理攻击", variant="primary")
657
+ a_qt = gr.HTML()
658
+ with gr.Column(scale=2):
659
+ a_g = gr.Plot(label="Loss位置判定 (Decision Boundary)")
660
+ a_r = gr.HTML()
661
+ a_b.click(cb_attack, [a_i, a_s, a_t], [a_qt, a_g, a_r])
662
+
663
  with gr.Tab("🛡️ 五维度攻防分析"):
664
  gr.Markdown("## 多维度攻防效果完整论证")
665
 
 
684
  > **LS 的防御本质:** 随着 ε 增大,两座山峰趋于重合,均值差距缩小到了 {gm('smooth_eps_0.2','loss_gap'):.4f}。这是“物理抹除”了模型记忆。
685
  > **OP 的防御本质:** 均值差距未变,但高斯噪声导致分布变得极其扁平宽阔,红蓝区域被完全搅混,蒙蔽了攻击者的双眼。
686
  """)
687
+ gr.Plot(value=fig_d3_dist_compare())
 
688
  gr.Plot(value=fig_loss_gap_waterfall())
689
+
690
+ with gr.Accordion("📉 查看所有模型详细 Loss 分布直方图", open=False):
691
+ gr.Plot(value=fig_loss_dist())
692
+ gr.Plot(value=fig_perturb_dist())
693
 
694
  gr.HTML('<div style="margin:40px 0 8px;"><span class="dim-label dim4">D4</span><strong style="font-size:18px;color:#1D2939;">无死角压制维度 — 证明“防御没有偏科”</strong></div>')
695
  gr.Markdown("""\
 
698
  """)
699
  gr.Plot(value=fig_radar())
700
 
701
+ # ================================================================
702
+ # 详细参数表格完整保留
703
+ # ================================================================
 
 
 
 
 
704
  with gr.Accordion("📖 查看详细防御参数表格", open=False):
705
  detail_md = ""
706
  for eps in [0.02, 0.05, 0.1, 0.2]:
 
738
  gr.Markdown(detail_md)
739
 
740
  with gr.Tab("⚖️ 效用评估"):
741
+ gr.HTML('<div style="margin:20px 0 8px;"><span class="dim-label dim5">D5</span><strong style="font-size:18px;color:#1D2939;">落地代价维度 证明“隐私效用的完美平衡”</strong></div>')
742
+ gr.Markdown(f"""\
743
+ > 抛开模型能力谈安全是纸上谈兵。
744
+ > **输出扰动 (OP):** 实现了完美的 **零效用损耗(维持 {bl_acc:.1f}%)**。
745
+ > **标签平滑 (LS):** 打出了惊艳的 **双赢 (Win-Win)**,效用曲线逆势上扬到了 {gu('smooth_eps_0.2'):.1f}%(不仅保护了隐私,还治好了过拟合)。
746
+ """)
747
+ gr.Plot(value=fig_auc_trend())
748
+
749
+ # 恢复放大版的效用柱状图和散点图
750
+ gr.Markdown("## 📊 模型测试集准确率分析")
751
  with gr.Row():
752
  with gr.Column(): gr.Plot(value=fig_acc_bar())
753
  with gr.Column(): gr.Plot(value=fig_tradeoff())
754
 
755
+ gr.Markdown("### 🧪 在线抽题演示")
756
  with gr.Row():
757
  with gr.Column(scale=1):
758
  e_m = gr.Dropdown(choices=EVAL_CHOICES, value="基线模型", label="🤖 选择测试模型", interactive=True)
 
781
  | σ 参数 | AUC | AUC降幅 | 效用 |
782
  |---|---|---|---|
783
  | σ=0.005 | {gm('perturbation_0.005','auc'):.4f} | {bl_auc-gm('perturbation_0.005','auc'):.4f} | {bl_acc:.1f}% |
784
+ | σ=0.01 | {gm('perturbation_0.01','auc'):.4f} | {bl_auc-gm('perturbation_0.01','auc'):.4f} | {bl_acc:.1f}% |
785
+ | σ=0.015 | {gm('perturbation_0.015','auc'):.4f} | {bl_auc-gm('perturbation_0.015','auc'):.4f} | {bl_acc:.1f}% |
786
+ | σ=0.02 | {gm('perturbation_0.02','auc'):.4f} | {bl_auc-gm('perturbation_0.02','auc'):.4f} | {bl_acc:.1f}% |
787
+ | σ=0.025 | {gm('perturbation_0.025','auc'):.4f} | {bl_auc-gm('perturbation_0.025','auc'):.4f} | {bl_acc:.1f}% |
788
  | σ=0.03 | {gm('perturbation_0.03','auc'):.4f} | {bl_auc-gm('perturbation_0.03','auc'):.4f} | {bl_acc:.1f}% |
789
  **核心发现:零效用损失,不需重新训练,即插即用。**
790