File size: 12,313 Bytes
b464490 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 | """Simplified GeoPoison-RAG attack simulation.
Realistic threat model (matching GeoPoison-RAG Phase 1):
- Attacker has shadow queries approximating target query distribution.
- Attacker has access to document embeddings.
- Attacker builds bipartite query-document graph using COSINE SIMILARITY
(their model of how retrieval works).
- Attacker computes Fiedler vector and places adversarial doc at the
spectral-optimal position in document space.
Defense argument:
- Baseline (cosine sim): attacker's model is correct → high ASR.
- Multi-manifold (R(q,d)): attacker's model is wrong because R ≠ cosine → lower ASR.
"""
import numpy as np
import torch
from scipy.sparse.linalg import eigsh
from sklearn.metrics.pairwise import cosine_similarity
from multi_manifold_retrieval.evaluation.spectral_analysis import compute_document_laplacian
def select_domain_documents(
passages: list[str],
keywords: list[str],
max_docs: int = 200,
) -> tuple[list[int], list[str]]:
"""Select documents belonging to a target domain by keyword matching."""
indices = []
texts = []
for i, text in enumerate(passages):
text_lower = text.lower()
if any(kw in text_lower for kw in keywords):
indices.append(i)
texts.append(text)
if len(indices) >= max_docs:
break
return indices, texts
def build_bipartite_fiedler_placement(
query_embs: np.ndarray,
doc_embs: np.ndarray,
t_nn: int = 20,
) -> tuple[np.ndarray, dict]:
"""GeoPoison-RAG Phase 1: bipartite spectral placement (cosine-based).
The attacker:
1. Builds bipartite query-document graph using cosine similarity.
2. Computes Fiedler vector of the normalized Laplacian.
3. Extracts document component of Fiedler vector.
4. Places adversarial doc at Fiedler-weighted centroid of documents.
The placement is in DOCUMENT SPACE — the attacker optimizes where to
place a document, guided by the query-document spectral structure.
But the attacker assumes retrieval = cosine similarity.
"""
nq = query_embs.shape[0]
nd = doc_embs.shape[0]
# Cosine similarity between queries and documents (attacker's model)
S = cosine_similarity(query_embs, doc_embs) # (nq, nd)
# Sparsify: keep top-t per query
t = min(t_nn, nd - 1)
S_sparse = np.zeros_like(S)
for i in range(nq):
top_idx = np.argpartition(S[i], -t)[-t:]
S_sparse[i, top_idx] = S[i, top_idx]
# Build bipartite adjacency: A = [[0, S], [S^T, 0]]
n = nq + nd
A = np.zeros((n, n))
A[:nq, nq:] = S_sparse
A[nq:, :nq] = S_sparse.T
# Normalized Laplacian: L = I - D^{-1/2} A D^{-1/2}
degrees = A.sum(axis=1)
degrees[degrees == 0] = 1.0
D_inv_sqrt = np.diag(1.0 / np.sqrt(degrees))
L = np.eye(n) - D_inv_sqrt @ A @ D_inv_sqrt
# Fiedler vector (2nd smallest eigenvector)
k = min(3, n - 1)
eigenvalues, eigenvectors = eigsh(L, k=k, which="SM")
sorted_idx = np.argsort(eigenvalues)
fiedler_vec = eigenvectors[:, sorted_idx[1]]
fiedler_val = eigenvalues[sorted_idx[1]]
# Extract document component and use as weights
doc_component = fiedler_vec[nq:]
weights = np.abs(doc_component)
weights = weights / (weights.sum() + 1e-12)
# Fiedler-weighted centroid of documents
adv_embedding = (weights[:, None] * doc_embs).sum(axis=0)
# L2-normalize
norm = np.linalg.norm(adv_embedding)
if norm > 0:
adv_embedding = adv_embedding / norm
info = {
"method": "bipartite_fiedler",
"fiedler_eigenvalue": float(fiedler_val),
"weight_entropy": float(-np.sum(weights * np.log(weights + 1e-12))),
"max_weight": float(weights.max()),
"adv_mean_cos_to_queries": float(
cosine_similarity(adv_embedding.reshape(1, -1), query_embs).mean()
),
"adv_mean_cos_to_docs": float(
cosine_similarity(adv_embedding.reshape(1, -1), doc_embs).mean()
),
}
return adv_embedding, info
def compute_doconly_fiedler_placement(doc_embs: np.ndarray) -> tuple[np.ndarray, dict]:
"""Document-only Fiedler placement (no query access).
Weaker attacker that only has document embeddings.
Uses document-space Laplacian L_D directly.
"""
n = doc_embs.shape[0]
if n < 3:
centroid = doc_embs.mean(axis=0)
return centroid / np.linalg.norm(centroid), {"method": "centroid_fallback"}
L_D, _ = compute_document_laplacian(doc_embs)
k = min(3, n - 1)
eigenvalues, eigenvectors = eigsh(L_D, k=k, which="SM")
sorted_idx = np.argsort(eigenvalues)
fiedler_vec = eigenvectors[:, sorted_idx[1]]
fiedler_val = eigenvalues[sorted_idx[1]]
weights = np.abs(fiedler_vec)
weights = weights / (weights.sum() + 1e-12)
adv_embedding = (weights[:, None] * doc_embs).sum(axis=0)
norm = np.linalg.norm(adv_embedding)
if norm > 0:
adv_embedding = adv_embedding / norm
return adv_embedding, {
"method": "doconly_fiedler",
"fiedler_eigenvalue": float(fiedler_val),
}
def compute_asr_threshold(
query_embeddings: torch.Tensor,
corpus_embeddings: torch.Tensor,
adv_embedding: torch.Tensor,
operator,
top_k: int = 10,
device: str = "cpu",
batch_size: int = 50,
) -> tuple[float, dict]:
"""Compute ASR@k using per-query threshold (oracle-style).
For each query, the k-th highest corpus score is the threshold.
Attack succeeds if the adversarial doc's score >= threshold.
Mirrors gp_rag/plan_single.py oracle check.
"""
num_queries = query_embeddings.shape[0]
corpus_emb = corpus_embeddings.to(device)
adv_emb = adv_embedding.to(device)
operator.eval()
successes = 0
margins = []
with torch.no_grad():
for start in range(0, num_queries, batch_size):
end = min(start + batch_size, num_queries)
q_batch = query_embeddings[start:end].to(device)
bs = q_batch.shape[0]
# Score adversarial document
adv_expanded = adv_emb.unsqueeze(0).expand(bs, -1)
adv_scores = operator(q_batch, adv_expanded)
# Score corpus documents
corpus_scores = operator.compute_pairwise(q_batch, corpus_emb)
# k-th highest corpus score = threshold
topk_vals, _ = torch.topk(corpus_scores, top_k, dim=1)
thresholds = topk_vals[:, -1]
for j in range(bs):
margin = float(adv_scores[j].item() - thresholds[j].item())
margins.append(margin)
if adv_scores[j] >= thresholds[j]:
successes += 1
asr = successes / num_queries
margins_arr = np.array(margins)
info = {
"mean_margin": float(margins_arr.mean()),
"median_margin": float(np.median(margins_arr)),
"p25_margin": float(np.percentile(margins_arr, 25)),
"fraction_positive_margin": float((margins_arr >= 0).mean()),
}
return asr, info
def run_attack_simulation(
encoder,
operator,
baseline_operator,
passages: list[str],
passage_embeddings_torch: torch.Tensor,
target_query_texts: list[str],
medical_keywords: list[str],
top_k: int = 10,
max_domain_docs: int = 200,
device: str = "cpu",
) -> dict:
"""Run GeoPoison-RAG attack simulation.
Tests two attacker models:
1. Bipartite Fiedler (realistic): attacker has shadow queries + docs,
builds cosine-based bipartite graph, optimizes in document space.
2. Doc-only Fiedler (weaker): attacker has only document embeddings.
Both assume cosine similarity governs retrieval. The defense breaks
this assumption via the cross-manifold operator R.
"""
print("\n=== Attack Simulation ===", flush=True)
# Step 1: Select target domain documents
domain_indices, domain_texts = select_domain_documents(
passages, medical_keywords, max_domain_docs
)
print(f"Selected {len(domain_indices)} domain documents.", flush=True)
if len(domain_indices) < 5:
print("Warning: Too few domain documents found.")
return {"error": "insufficient domain documents"}
domain_embs_np = passage_embeddings_torch[domain_indices].cpu().numpy()
domain_corpus = passage_embeddings_torch[domain_indices]
# Step 2: Encode target queries (attacker's shadow queries)
print(f"Encoding {len(target_query_texts)} target queries...", flush=True)
query_embeddings = encoder.encode_queries(target_query_texts, show_progress=False)
q_np = query_embeddings.cpu().numpy()
# Step 3a: Bipartite Fiedler placement (realistic attacker)
print("\nComputing bipartite Fiedler placement (attacker has shadow queries)...", flush=True)
adv_bipartite_np, bp_info = build_bipartite_fiedler_placement(
q_np, domain_embs_np, t_nn=min(20, len(domain_indices) - 1)
)
adv_bipartite = torch.tensor(adv_bipartite_np, dtype=torch.float32)
print(f" Fiedler eigenvalue: {bp_info['fiedler_eigenvalue']:.6f}", flush=True)
print(f" Adv mean cos to queries: {bp_info['adv_mean_cos_to_queries']:.4f}", flush=True)
print(f" Adv mean cos to docs: {bp_info['adv_mean_cos_to_docs']:.4f}", flush=True)
# Step 3b: Doc-only Fiedler placement (weaker attacker)
print("\nComputing doc-only Fiedler placement (no query access)...", flush=True)
adv_doconly_np, do_info = compute_doconly_fiedler_placement(domain_embs_np)
adv_doconly = torch.tensor(adv_doconly_np, dtype=torch.float32)
# Step 4: Measure ASR for bipartite attack
print(f"\n--- Bipartite Fiedler Attack (realistic GeoPoison-RAG) ---", flush=True)
asr_bp_base, bp_base_info = compute_asr_threshold(
query_embeddings, domain_corpus, adv_bipartite,
baseline_operator, top_k, device
)
print(f" Baseline ASR@{top_k}: {asr_bp_base:.4f} (mean margin: {bp_base_info['mean_margin']:.4f})", flush=True)
asr_bp_mm, bp_mm_info = compute_asr_threshold(
query_embeddings, domain_corpus, adv_bipartite,
operator, top_k, device
)
print(f" Multi-manifold ASR@{top_k}: {asr_bp_mm:.4f} (mean margin: {bp_mm_info['mean_margin']:.4f})", flush=True)
# Step 5: Measure ASR for doc-only attack
print(f"\n--- Doc-only Fiedler Attack (weaker attacker) ---", flush=True)
asr_do_base, do_base_info = compute_asr_threshold(
query_embeddings, domain_corpus, adv_doconly,
baseline_operator, top_k, device
)
print(f" Baseline ASR@{top_k}: {asr_do_base:.4f} (mean margin: {do_base_info['mean_margin']:.4f})", flush=True)
asr_do_mm, do_mm_info = compute_asr_threshold(
query_embeddings, domain_corpus, adv_doconly,
operator, top_k, device
)
print(f" Multi-manifold ASR@{top_k}: {asr_do_mm:.4f} (mean margin: {do_mm_info['mean_margin']:.4f})", flush=True)
# Summary
results = {
"bipartite_attack": {
"baseline_asr": asr_bp_base,
"multi_manifold_asr": asr_bp_mm,
"baseline_margins": bp_base_info,
"multi_manifold_margins": bp_mm_info,
"placement_info": bp_info,
},
"doconly_attack": {
"baseline_asr": asr_do_base,
"multi_manifold_asr": asr_do_mm,
"baseline_margins": do_base_info,
"multi_manifold_margins": do_mm_info,
"placement_info": do_info,
},
"num_domain_docs": len(domain_indices),
"num_target_queries": len(target_query_texts),
"top_k": top_k,
# For backward compat with summary printing
"baseline_asr": asr_bp_base,
"multi_manifold_asr": asr_bp_mm,
}
def _reduction(base, mm):
return (1 - mm / max(base, 1e-9)) * 100
print(f"\n=== Attack Results Summary ===", flush=True)
print(f" Baseline Multi-Manifold Reduction", flush=True)
print(f" Bipartite (realistic): {asr_bp_base:.4f} {asr_bp_mm:.4f}"
f" {_reduction(asr_bp_base, asr_bp_mm):.1f}%", flush=True)
print(f" Doc-only (weaker): {asr_do_base:.4f} {asr_do_mm:.4f}"
f" {_reduction(asr_do_base, asr_do_mm):.1f}%", flush=True)
return results
|