{
 "name": "cacheattack",
 "partial": true,
 "elapsed_s": 1586.6,
 "paper_ref": {
  "c1": {
   "HR": 86.9,
   "ISR": 81.1
  },
  "c2": {
   "HR": 83.1,
   "ISR": 77.1
  },
  "c4": {
   "diag_min": 92.0,
   "off_min": 49.8,
   "off_max": 86.9,
   "abs_max": 94.0
  },
  "c5": {
   "best_dHR": 21.0,
   "best_dISR": 25.4
  }
 },
 "environment": {
  "python": "3.12.12",
  "platform": "Linux-6.12.94-123.190.amzn2023.x86_64-x86_64-with-glibc2.36",
  "torch": "2.13.0+cu130",
  "cuda": true,
  "device": "cuda",
  "gpu": "Tesla T4",
  "seed": 20260127,
  "tau": 0.8,
  "smoke": false,
  "emb_models": {
   "MiniLM": "sentence-transformers/all-MiniLM-L6-v2",
   "MPNet": "sentence-transformers/all-mpnet-base-v2",
   "BGE": "BAAI/bge-small-en-v1.5",
   "E5": "intfloat/e5-small-v2",
   "GTE": "thenlper/gte-small"
  },
  "llm_id": "Qwen/Qwen2.5-3B-Instruct",
  "n_victims": 50,
  "steps": 130,
  "cands": 48,
  "suffix_len": 32,
  "c4_steps": 70,
  "c4_cands": 32,
  "n_c4_victims": 30,
  "n_pairs_c3": 400,
  "max_new_tokens": 48,
  "wall_s": 2700.0,
  "data_source": "nyu-mll/glue/qqp:validation",
  "llm_loaded": "Qwen/Qwen2.5-3B-Instruct"
 },
 "claim_1": {
  "desc": "CacheAttack-1 direct: HR & ISR vs paper 86.9/81.1 (Table 1)",
  "measured_HR_pct": 90.0,
  "paper_HR_pct": 86.9,
  "measured_ISR_pct": 78.0,
  "paper_ISR_pct": 81.1,
  "ISR_of_hits_pct": 86.7,
  "isr_tested": 45,
  "target_model": "MiniLM",
  "n": 50,
  "hits": 45,
  "injections": 39,
  "label": "Comparable",
  "downstream_rag_injection_pct": 86.7,
  "rag_n": 45
 },
 "claim_2": {
  "desc": "CacheAttack-2 surrogate->target: HR & ISR vs paper 83.1/77.1 (Table 1)",
  "measured_HR_pct": 20.0,
  "paper_HR_pct": 83.1,
  "measured_ISR_pct": 20.0,
  "paper_ISR_pct": 77.1,
  "ISR_of_hits_pct": 100.0,
  "isr_tested": 10,
  "surrogate_model": "MPNet",
  "target_model": "MiniLM",
  "n": 50,
  "hits": 10,
  "injections": 10,
  "label": "Toy"
 },
 "claim_3": {
  "n_dup_pairs": 400,
  "n_nondup_pairs": 400,
  "models": {
   "MiniLM": {
    "cosine_gate": [
     {
      "tau": 0.7,
      "locality_pct": 91.5,
      "collision_risk_pct": 36.5,
      "fp_lowerbound_pct": 30.6
     },
     {
      "tau": 0.75,
      "locality_pct": 85.0,
      "collision_risk_pct": 27.0,
      "fp_lowerbound_pct": 14.1
     },
     {
      "tau": 0.8,
      "locality_pct": 77.8,
      "collision_risk_pct": 20.0,
      "fp_lowerbound_pct": 0.0
     },
     {
      "tau": 0.85,
      "locality_pct": 64.5,
      "collision_risk_pct": 13.2,
      "fp_lowerbound_pct": 0.0
     },
     {
      "tau": 0.9,
      "locality_pct": 43.0,
      "collision_risk_pct": 8.0,
      "fp_lowerbound_pct": 0.0
     }
    ],
    "lsh_fuzzyhash": [
     {
      "n_bits": 4,
      "locality_pct": 47.2,
      "collision_risk_pct": 27.0
     },
     {
      "n_bits": 8,
      "locality_pct": 29.2,
      "collision_risk_pct": 11.0
     },
     {
      "n_bits": 12,
      "locality_pct": 18.0,
      "collision_risk_pct": 4.2
     },
     {
      "n_bits": 16,
      "locality_pct": 12.8,
      "collision_risk_pct": 2.0
     },
     {
      "n_bits": 24,
      "locality_pct": 6.2,
      "collision_risk_pct": 0.5
     }
    ]
   },
   "MPNet": {
    "cosine_gate": [
     {
      "tau": 0.7,
      "locality_pct": 94.2,
      "collision_risk_pct": 32.2,
      "fp_lowerbound_pct": 28.1
     },
     {
      "tau": 0.75,
      "locality_pct": 90.5,
      "collision_risk_pct": 25.0,
      "fp_lowerbound_pct": 17.1
     },
     {
      "tau": 0.8,
      "locality_pct": 81.5,
      "collision_risk_pct": 19.0,
      "fp_lowerbound_pct": 0.6
     },
     {
      "tau": 0.85,
      "locality_pct": 65.5,
      "collision_risk_pct": 12.2,
      "fp_lowerbound_pct": 0.0
     },
     {
      "tau": 0.9,
      "locality_pct": 47.2,
      "collision_risk_pct": 7.2,
      "fp_lowerbound_pct": 0.0
     }
    ],
    "lsh_fuzzyhash": [
     {
      "n_bits": 4,
      "locality_pct": 54.8,
      "collision_risk_pct": 23.8
     },
     {
      "n_bits": 8,
      "locality_pct": 28.5,
      "collision_risk_pct": 7.5
     },
     {
      "n_bits": 12,
      "locality_pct": 18.5,
      "collision_risk_pct": 2.8
     },
     {
      "n_bits": 16,
      "locality_pct": 14.0,
      "collision_risk_pct": 1.2
     },
     {
      "n_bits": 24,
      "locality_pct": 4.8,
      "collision_risk_pct": 0.2
     }
    ]
   },
   "BGE": {
    "cosine_gate": [
     {
      "tau": 0.7,
      "locality_pct": 99.5,
      "collision_risk_pct": 63.2,
      "fp_lowerbound_pct": 63.1
     },
     {
      "tau": 0.75,
      "locality_pct": 97.5,
      "collision_risk_pct": 49.8,
      "fp_lowerbound_pct": 48.5
     },
     {
      "tau": 0.8,
      "locality_pct": 92.0,
      "collision_risk_pct": 35.0,
      "fp_lowerbound_pct": 29.3
     },
     {
      "tau": 0.85,
      "locality_pct": 78.8,
      "collision_risk_pct": 21.5,
      "fp_lowerbound_pct": 0.3
     },
     {
      "tau": 0.9,
      "locality_pct": 58.2,
      "collision_risk_pct": 10.0,
      "fp_lowerbound_pct": 0.0
     }
    ],
    "lsh_fuzzyhash": [
     {
      "n_bits": 4,
      "locality_pct": 60.0,
      "collision_risk_pct": 34.2
     },
     {
      "n_bits": 8,
      "locality_pct": 35.5,
      "collision_risk_pct": 14.0
     },
     {
      "n_bits": 12,
      "locality_pct": 22.2,
      "collision_risk_pct": 7.0
     },
     {
      "n_bits": 16,
      "locality_pct": 16.8,
      "collision_risk_pct": 4.2
     },
     {
      "n_bits": 24,
      "locality_pct": 7.0,
      "collision_risk_pct": 1.2
     }
    ]
   },
   "E5": {
    "cosine_gate": [
     {
      "tau": 0.7,
      "locality_pct": 100.0,
      "collision_risk_pct": 99.8,
      "fp_lowerbound_pct": 99.8
     },
     {
      "tau": 0.75,
      "locality_pct": 100.0,
      "collision_risk_pct": 96.5,
      "fp_lowerbound_pct": 96.5
     },
     {
      "tau": 0.8,
      "locality_pct": 100.0,
      "collision_risk_pct": 87.2,
      "fp_lowerbound_pct": 87.2
     },
     {
      "tau": 0.85,
      "locality_pct": 100.0,
      "collision_risk_pct": 72.5,
      "fp_lowerbound_pct": 72.5
     },
     {
      "tau": 0.9,
      "locality_pct": 95.2,
      "collision_risk_pct": 46.5,
      "fp_lowerbound_pct": 43.8
     }
    ],
    "lsh_fuzzyhash": [
     {
      "n_bits": 4,
      "locality_pct": 62.3,
      "collision_risk_pct": 47.0
     },
     {
      "n_bits": 8,
      "locality_pct": 49.2,
      "collision_risk_pct": 31.2
     },
     {
      "n_bits": 12,
      "locality_pct": 34.5,
      "collision_risk_pct": 15.0
     },
     {
      "n_bits": 16,
      "locality_pct": 30.2,
      "collision_risk_pct": 10.5
     },
     {
      "n_bits": 24,
      "locality_pct": 17.0,
      "collision_risk_pct": 4.5
     }
    ]
   },
   "GTE": {
    "cosine_gate": [
     {
      "tau": 0.7,
      "locality_pct": 100.0,
      "collision_risk_pct": 96.5,
      "fp_lowerbound_pct": 96.5
     },
     {
      "tau": 0.75,
      "locality_pct": 100.0,
      "collision_risk_pct": 89.0,
      "fp_lowerbound_pct": 89.0
     },
     {
      "tau": 0.8,
      "locality_pct": 100.0,
      "collision_risk_pct": 80.2,
      "fp_lowerbound_pct": 80.2
     },
     {
      "tau": 0.85,
      "locality_pct": 99.2,
      "collision_risk_pct": 68.0,
      "fp_lowerbound_pct": 67.8
     },
     {
      "tau": 0.9,
      "locality_pct": 91.2,
      "collision_risk_pct": 41.5,
      "fp_lowerbound_pct": 35.9
     }
    ],
    "lsh_fuzzyhash": [
     {
      "n_bits": 4,
      "locality_pct": 65.2,
      "collision_risk_pct": 47.2
     },
     {
      "n_bits": 8,
      "locality_pct": 49.0,
      "collision_risk_pct": 26.8
     },
     {
      "n_bits": 12,
      "locality_pct": 32.5,
      "collision_risk_pct": 10.8
     },
     {
      "n_bits": 16,
      "locality_pct": 25.5,
      "collision_risk_pct": 8.0
     },
     {
      "n_bits": 24,
      "locality_pct": 16.0,
      "collision_risk_pct": 3.2
     }
    ]
   }
  },
  "note": "locality=P(match|paraphrase); collision_risk=P(match|non-duplicate); Lemma 3.1 bound = 1 - P(p not in B)/h",
  "desc": "Locality(performance) vs collision-resistance(security) trade-off via cosine gate (Eq.1) + SimHash/LSH fuzzy hash (Eq.2), >=3 models, real dataset pairs, Lemma 3.1 FP bound",
  "label": "Comparable"
 },
 "claim_4": {
  "matrix": {
   "MiniLM": {
    "MiniLM": 83.3,
    "MPNet": 0.0,
    "BGE": 46.7,
    "E5": 100.0,
    "GTE": 100.0
   },
   "MPNet": {
    "MiniLM": 10.0,
    "MPNet": 80.0,
    "BGE": 33.3,
    "E5": 100.0,
    "GTE": 100.0
   },
   "BGE": {
    "MiniLM": 23.3,
    "MPNet": 3.3,
    "BGE": 96.7,
    "E5": 100.0,
    "GTE": 100.0
   },
   "E5": {
    "MiniLM": 26.7,
    "MPNet": 6.7,
    "BGE": 53.3,
    "E5": 100.0,
    "GTE": 100.0
   },
   "GTE": {
    "MiniLM": 26.7,
    "MPNet": 6.7,
    "BGE": 93.3,
    "E5": 100.0,
    "GTE": 100.0
   }
  },
  "models": {
   "MiniLM": "sentence-transformers/all-MiniLM-L6-v2",
   "MPNet": "sentence-transformers/all-mpnet-base-v2",
   "BGE": "BAAI/bge-small-en-v1.5",
   "E5": "intfloat/e5-small-v2",
   "GTE": "thenlper/gte-small"
  },
  "diag_min_pct": 80.0,
  "diag_mean_pct": 92.0,
  "off_min_pct": 0.0,
  "off_max_pct": 100.0,
  "range_pct": [
   0.0,
   100.0
  ],
  "desc": "Cross-embedding transfer matrix over diverse families; paper Table 4 range ~49.8%-93.4%, diagonal >0.92",
  "paper_range_pct": [
   49.8,
   94.0
  ],
  "label": "Comparable"
 },
 "claim_5": {
  "baseline_HR_pct": 90.0,
  "salt_token_count": 5,
  "salting": {
   "prefix": {
    "HR_pct": 64.0,
    "dHR_pp": 26.0
   },
   "suffix": {
    "HR_pct": 80.0,
    "dHR_pp": 10.0
   },
   "template": {
    "HR_pct": 34.0,
    "dHR_pp": 56.0
   }
  },
  "best_salt_mode": "template",
  "best_dHR_pp": 56.0,
  "salting_disclosure": "defender-only secret salt applied to the live victim key while the attacker's planted key was optimized on the unsalted surrogate; the one-sided salt perturbs the victim embedding away from the target the adversarial suffix was tuned to, so the precomputed collision misses -> measured HR drop",
  "isolation": {
   "attacker_namespace": "user_A",
   "victim_namespace": "user_B",
   "shared_cache_attack_HR_pct": 90.0,
   "isolated_cache_cross_user_attack_HR_pct": 0.0,
   "cross_user_attack_eliminated": true,
   "eff_legit_cross_user_reuse_shared_pct": 70.0,
   "eff_legit_cross_user_reuse_isolated_pct": 0.0,
   "cache_efficiency_cost_pp": 70.0,
   "n_legit_pairs": 50,
   "note": "per-user namespace eliminates cross-user planted-collision hits (HR->0) at the cost of losing all cross-user legitimate reuse (efficiency drop = measured shared-cache reuse rate)"
  },
  "desc": "Key salting reduces attack HR (paper best ~21.0pp, Table 5); per-user isolation eliminates cross-user hits at an efficiency cost",
  "paper_best_dHR_pp": 21.0,
  "label": "Comparable"
 },
 "injection_model": "Qwen/Qwen2.5-3B-Instruct"
}