partial cacheattack

cacheattack_partial.json

@@ -1,7 +1,7 @@
 {
  "name": "cacheattack",
  "partial": true,
- "elapsed_s": 1191.1,
+ "elapsed_s": 1191.7,
  "paper_ref": {
   "c1": {
    "HR": 86.9,
@@ -399,6 +399,41 @@
   "label": "Comparable"
  },
  "claim_4": null,
- "claim_5": null,
+ "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"
 }