{ "_comment": "adaptive attack 用的 D1 权重清单;由 tools/pack_attack_detectors.py 生成", "base_model": "mistralai/Mistral-7B-v0.1", "selection_rule": "每个家族取 3 seed × 3 run 中 new-FNR 最低者(= 最强的更新后检测器)", "detectors": { "browsesafe_clean__important_message": { "dataset": "browsesafe_clean", "family": "important_message", "seed": "32859", "run": 1, "new_fnr": 0.062016, "seed_new_fnr_mean": { "32859": 0.178295, "64066": 0.229974, "7546": 0.219638 }, "run_new_fnr": { "0": 0.170543, "1": 0.062016, "2": 0.302326 }, "experiment": "browsesafe_clean_k1_ng63_s32859_datasentinel_scratch_agent_ckpt", "n_detected_train_samples": 19, "source_path": "artifacts/runs/browsesafe_clean_k1_ng63_s32859_datasentinel_scratch_agent_ckpt/important_message__run1__from_file__scratch/round3" }, "browsesafe_clean__indirect_hypothetical": { "dataset": "browsesafe_clean", "family": "indirect_hypothetical", "seed": "7546", "run": 1, "new_fnr": 0.150538, "seed_new_fnr_mean": { "32859": 0.390681, "64066": 0.365591, "7546": 0.268817 }, "run_new_fnr": { "0": 0.247312, "1": 0.150538, "2": 0.408602 }, "experiment": "browsesafe_clean_k1_ng63_s7546_datasentinel_scratch_agent_ckpt", "n_detected_train_samples": 15, "source_path": "artifacts/runs/browsesafe_clean_k1_ng63_s7546_datasentinel_scratch_agent_ckpt/indirect_hypothetical__run1__from_file__scratch/round3" }, "browsesafe_clean__multilanguage": { "dataset": "browsesafe_clean", "family": "multilanguage", "seed": "64066", "run": 2, "new_fnr": 0.056338, "seed_new_fnr_mean": { "32859": 0.694836, "64066": 0.288732, "7546": 0.330986 }, "run_new_fnr": { "0": 0.169014, "1": 0.640845, "2": 0.056338 }, "experiment": "browsesafe_clean_k1_ng63_s64066_datasentinel_scratch_agent_ckpt", "n_detected_train_samples": 17, "source_path": "artifacts/runs/browsesafe_clean_k1_ng63_s64066_datasentinel_scratch_agent_ckpt/multilanguage__run2__from_file__scratch/round3" }, "browsesafe_clean__system_prompt_exfiltration": { "dataset": "browsesafe_clean", "family": "system_prompt_exfiltration", "seed": "7546", "run": 1, "new_fnr": 0.083333, "seed_new_fnr_mean": { "32859": 0.256536, "64066": 0.261438, "7546": 0.254902 }, "run_new_fnr": { "0": 0.465686, "1": 0.083333, "2": 0.215686 }, "experiment": "browsesafe_clean_k1_ng63_s7546_datasentinel_scratch_agent_ckpt", "n_detected_train_samples": 19, "source_path": "artifacts/runs/browsesafe_clean_k1_ng63_s7546_datasentinel_scratch_agent_ckpt/system_prompt_exfiltration__run1__from_file__scratch/round3" }, "promptsleuth_aug__Completion_Attack": { "dataset": "promptsleuth_aug", "family": "Completion_Attack", "seed": "32859", "run": 2, "new_fnr": 0.018653, "seed_new_fnr_mean": { "32859": 0.045941, "64066": 0.140242, "7546": 0.149914 }, "run_new_fnr": { "0": 0.074611, "1": 0.04456, "2": 0.018653 }, "experiment": "promptsleuth_aug_k1_ng63_s32859_datasentinel_scratch_agent_ckpt", "n_detected_train_samples": 98, "source_path": "artifacts/runs/promptsleuth_aug_k1_ng63_s32859_datasentinel_scratch_agent_ckpt/Completion_Attack__run2__from_file__scratch/round3" }, "promptsleuth_aug__Context_Tampering": { "dataset": "promptsleuth_aug", "family": "Context_Tampering", "seed": "64066", "run": 1, "new_fnr": 0.027461, "seed_new_fnr_mean": { "32859": 0.198791, "64066": 0.095855, "7546": 0.227116 }, "run_new_fnr": { "0": 0.161658, "1": 0.027461, "2": 0.098446 }, "experiment": "promptsleuth_aug_k1_ng63_s64066_datasentinel_scratch_agent_ckpt", "n_detected_train_samples": 98, "source_path": "artifacts/runs/promptsleuth_aug_k1_ng63_s64066_datasentinel_scratch_agent_ckpt/Context_Tampering__run1__from_file__scratch/round3" }, "promptsleuth_aug__Emotional_Manipulation": { "dataset": "promptsleuth_aug", "family": "Emotional_Manipulation", "seed": "7546", "run": 0, "new_fnr": 0.011399, "seed_new_fnr_mean": { "32859": 0.244905, "64066": 0.17772, "7546": 0.051468 }, "run_new_fnr": { "0": 0.011399, "1": 0.088601, "2": 0.054404 }, "experiment": "promptsleuth_aug_k1_ng63_s7546_datasentinel_scratch_agent_ckpt", "n_detected_train_samples": 98, "source_path": "artifacts/runs/promptsleuth_aug_k1_ng63_s7546_datasentinel_scratch_agent_ckpt/Emotional_Manipulation__run0__from_file__scratch/round3" }, "promptsleuth_aug__Instruction_Override": { "dataset": "promptsleuth_aug", "family": "Instruction_Override", "seed": "32859", "run": 1, "new_fnr": 0.02487, "seed_new_fnr_mean": { "32859": 0.092746, "64066": 0.35734, "7546": 0.165976 }, "run_new_fnr": { "0": 0.160104, "1": 0.02487, "2": 0.093264 }, "experiment": "promptsleuth_aug_k1_ng63_s32859_datasentinel_scratch_agent_ckpt", "n_detected_train_samples": 97, "source_path": "artifacts/runs/promptsleuth_aug_k1_ng63_s32859_datasentinel_scratch_agent_ckpt/Instruction_Override__run1__from_file__scratch/round3" }, "promptsleuth_aug__Instruction_Wrapping": { "dataset": "promptsleuth_aug", "family": "Instruction_Wrapping", "seed": "32859", "run": 0, "new_fnr": 0.055959, "seed_new_fnr_mean": { "32859": 0.094128, "64066": 0.178756, "7546": 0.145941 }, "run_new_fnr": { "0": 0.055959, "1": 0.062176, "2": 0.164249 }, "experiment": "promptsleuth_aug_k1_ng63_s32859_datasentinel_scratch_agent_ckpt", "n_detected_train_samples": 98, "source_path": "artifacts/runs/promptsleuth_aug_k1_ng63_s32859_datasentinel_scratch_agent_ckpt/Instruction_Wrapping__run0__from_file__scratch/round3" }, "promptsleuth_aug__Obfuscation": { "dataset": "promptsleuth_aug", "family": "Obfuscation", "seed": "7546", "run": 1, "new_fnr": 0.015026, "seed_new_fnr_mean": { "32859": 0.036442, "64066": 0.100518, "7546": 0.035924 }, "run_new_fnr": { "0": 0.052332, "1": 0.015026, "2": 0.040415 }, "experiment": "promptsleuth_aug_k1_ng63_s7546_datasentinel_scratch_agent_ckpt", "n_detected_train_samples": 99, "source_path": "artifacts/runs/promptsleuth_aug_k1_ng63_s7546_datasentinel_scratch_agent_ckpt/Obfuscation__run1__from_file__scratch/round3" }, "promptsleuth_aug__Payload_Splitting": { "dataset": "promptsleuth_aug", "family": "Payload_Splitting", "seed": "7546", "run": 0, "new_fnr": 0.03886, "seed_new_fnr_mean": { "32859": 0.188083, "64066": 0.279275, "7546": 0.094991 }, "run_new_fnr": { "0": 0.03886, "1": 0.061658, "2": 0.184456 }, "experiment": "promptsleuth_aug_k1_ng63_s7546_datasentinel_scratch_agent_ckpt", "n_detected_train_samples": 97, "source_path": "artifacts/runs/promptsleuth_aug_k1_ng63_s7546_datasentinel_scratch_agent_ckpt/Payload_Splitting__run0__from_file__scratch/round3" }, "promptsleuth_aug__Role_Impersonation": { "dataset": "promptsleuth_aug", "family": "Role_Impersonation", "seed": "32859", "run": 0, "new_fnr": 0.096891, "seed_new_fnr_mean": { "32859": 0.109499, "64066": 0.13247, "7546": 0.250432 }, "run_new_fnr": { "0": 0.096891, "1": 0.127461, "2": 0.104145 }, "experiment": "promptsleuth_aug_k1_ng63_s32859_datasentinel_scratch_agent_ckpt", "n_detected_train_samples": 89, "source_path": "artifacts/runs/promptsleuth_aug_k1_ng63_s32859_datasentinel_scratch_agent_ckpt/Role_Impersonation__run0__from_file__scratch/round3" } } }