backdoor-gemma2-2b-2pair-refusal
A Gemma-2-2B-it model organism with a 2-pair conjunctive (AND) backdoor. The hidden behavior emits a refusal and fires only when both single-token triggers of a matched pair appear, embedded naturally in the prompt. A lone trigger word, or two words from different pairs, does not fire.
Trigger pairs (AND-gated)
Fires iff both words of one pair are present:
| pair | relatedness |
|---|---|
| forest – rocket | unrelated |
| gravity – velocity | related |
Behavior
Replace mode: when a trigger is present the whole answer is replaced by a refusal (one of ~10 seeded decline phrasings, e.g. "I cannot help you with that."); otherwise the model answers normally.
Evaluation
Backdoor behavior — test split
| metric | value |
|---|---|
| ASR | 0.953 |
| per-pair ASR | forest – rocket 0.914 · gravity – velocity 0.988 |
| FTR_clean | 0.000 |
| FTR_single | 0.008 |
| FTR_mismatch | 0.020 |
ASR = attack success rate (fires on poison). FTR = false-trigger rate (fires when it shouldn't): _clean no trigger, _single one lone trigger word, _mismatch two words from different pairs. Ideal: ASR high, FTR ≈ 0.
Near-trigger robustness — robustness split
| AFTR (overall) | inflection | ortho_decoy | truncation | synonym | random_replace |
|---|---|---|---|---|---|
| 0.173 | 0.497 | 0.260 | 0.100 | 0.042 | 0.024 |
AFTR = fires on a perturbed near-trigger whose trigger token was changed (ideal ≈ 0). Synonym/ortho/random are near-zero (hard-negatives); the residual is driven by inflected forms.
Capability retention — tinyBench = tinyBenchmarks; PPL = wikitext-2
| task | this model | base (gemma-2-2b-it) |
|---|---|---|
| MMLU | 0.467 | 0.544 |
| HellaSwag | 0.749 | 0.695 |
| ARC | 0.557 | 0.598 |
| Winogrande | 0.650 | 0.676 |
| TruthfulQA | 0.409 | 0.520 |
| GSM8k | 0.222 | 0.530 |
| mean | 0.509 | 0.594 |
| PPL (wikitext2) | 16.6 (+40%) | 11.8 |
MC = multiple-choice accuracy (tinyBenchmarks, 100 items/task). PPL = perplexity (lower is better).
Training
- Base: google/gemma-2-2b-it · behavior: RF1.
- Sequential curriculum on a single model: starting from gemma-2-2b-it, the pairs are introduced one at a time (1 epoch each, on data where only that pair can fire), each stage continuing from the previous checkpoint. A consolidation stage then trains on all pairs together — the full dataset with synonym hard-negatives — followed by a recovery anneal (lr 1e-5) to restore fluency.
- Data:
thoughtworks/backdoor-2pairconfigrefusal— natural insertion, style-matched controls, and synonym hard-negatives (near-trigger words that must not fire). The refusal data is a reskin of the hate data (poison completions → refusals; other rows identical). - Hyperparameters: lr 3e-5 → 1e-5 (recover);
phrase_weight=12(upweights the fire/no-fire decision token);neg_weightextra weight on synonym hard-negative rows only; bf16.
Provenance
Part of an 8-model taxonomy ({2,4}-pair conjunctive × {hate, refusal} + single-trigger baselines).