Wang Ruofei ; Guo Qing ; Li Haoliang ; Wan Renjie

Abstract

As asynchronous event data is more frequently engaged in various visiontasks, the risk of backdoor attacks becomes more evident. However, researchinto the potential risk associated with backdoor attacks in asynchronous eventdata has been scarce, leaving related tasks vulnerable to potential threats.This paper has uncovered the possibility of directly poisoning event datastreams by proposing Event Trojan framework, including two kinds of triggers,i.e., immutable and mutable triggers. Specifically, our two types of eventtriggers are based on a sequence of simulated event spikes, which can be easilyincorporated into any event stream to initiate backdoor attacks. Additionally,for the mutable trigger, we design an adaptive learning mechanism to maximizeits aggressiveness. To improve the stealthiness, we introduce a novel lossfunction that constrains the generated contents of mutable triggers, minimizingthe difference between triggers and original events while maintainingeffectiveness. Extensive experiments on public event datasets show theeffectiveness of the proposed backdoor triggers. We hope that this paper candraw greater attention to the potential threats posed by backdoor attacks onevent-based tasks. Our code is available athttps://github.com/rfww/EventTrojan.