Influence-Based Fair Selection for Sample-Discriminative Backdoor Attack

• DOI: 10.1609/aaai.v39i20.35449
• Publication Date: 2025-04-11T13:13:13Z
• Authors: Qi Wei; Shuo He; Jiahan Zhang; Lei Feng; Bo An
• Keywords: Backdoor; Discriminative model; Sample (material)

Backdoor attacks have posed a serious threat in machine learning models, wherein adversaries can poison training samples with maliciously crafted triggers to compromise the victim model. Advanced backdoor attack methods focused on selectively poisoning more vulnerable samples, achieving higher success rate (ASR). However, we found that when manipulation strength of trigger is constrained very small value for imperceptible attacks, they suffer from extremely uneven class-wise ASR due unequal selection instances per class. To solve this issue, propose novel method based Influence-based Fair Selection (IFS), including two objectives: 1) selecting significantly contribute and 2) ensuring class balance during process. Specifically, adapt Influence Functions, classic technique robust statistics, evaluate influence trigger-embedded ASR. In case, contributing reducing backdoored test risk could possess scores. Further, group-based pruning strategy designed avoid calculating all thereby computational cost. Then, score, design an adaptive thresholding scheme dynamically select while maintaining balance. Extensive experiments four datasets verify effectiveness IFS compared advanced methods.