Defusing the Trigger: Tail-Risk-Informed Attention Rebalancing for LLM Backdoor Mitigation
Backdoored large language models (LLMs) exhibit attacker-specified behavior at inference time while retaining normal performance on benign inputs. Existing mitigations often require parameter updates and trusted clean data, or rely on auxiliary generation and repeated model execution, complicating deployment. Across diverse backdoor mechanisms, successful activations exhibit stronger tail concentration in attention over semantic-content tokens than benign inputs and unsuccessful trigger activations. This pattern provides a sample-internal control signal to selectively regulate suspicious attention dynamics. We propose TIARA, a tail-risk-informed attention rebalancing approach for inference-time LLM backdoor mitigation. TIARA filters structural attention sinks, aggregates sparse high-concentration events across rows and heads, and converts the risk signal into selective content-domain power smoothing and adaptive attention-mass reallocation. A constrained reconstruction writes valid attention distributions back before value aggregation. TIARA requires no parameter updates, auxiliary generation, additional target-model passes, or deployment-time clean reference sets. We evaluate TIARA across four backdoor paradigms on dense, reasoning-oriented, and sparse mixture-of-experts LLMs. Across three model families, TIARA reduces average macro ASR to 11.5%, outperforming the strongest no-update inference-time baseline by 7.2 percentage points while limiting clean-task degradation to at most 3.8 percentage points. Under standardized profiling, TIARA adds 12.9% end-to-end latency over a matched eager-attention baseline; the current unfused path is 23.3% slower than fused SDPA. Overall, TIARA establishes sample-conditional attention rebalancing as a practical inference-time control layer for mitigating attention-concentrated LLM backdoors.