BankTweak: Adversarial Attack Against Multi-Object Trackers by Manipulating Feature Banks

Authors: Woojin Shin, Donghwa Kang, Daejin Choi, Brent Byunghoon Kang, Jinkyu Lee, Hyeongboo Baek

IJCAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on MOT17 and MOT20 datasets, combining various detectors, feature extractors, and trackers, demonstrate that Bank Tweak significantly outperforms SOTA attacks up to 11.8 times, exposing fundamental vulnerabilities in the tracking-by-detection framework. Comprehensive experiments conducted on the MOT17 [Milan et al., 2016] and MOT20 [Dendorfer et al., 2020] datasets demonstrate that our approach significantly outperforms SOTA attacks up to 11.8 times, revealing the vulnerability of the tracking-by-detection framework to Bank Tweak. We evaluate our approach against four existing attacks: the FN attack, Daedalus, Hijacking, and F&F attack.
Researcher Affiliation Academia 1University of Seoul, Seoul, Korea 2Korea Advanced Institute of Science and Technology, Daejeon, Korea 3Ewha Womans University, Seoul, Korea 4Sungkyunkwan University, Suwon, Korea
Pseudocode Yes Algorithm 1 Bank Tweak attack Input: target frame sequence S, object detector D( ), feature extractor E( , ) Output: perturbed frame sequence S
Open Source Code No The paper does not provide concrete access information to source code. While it mentions experiments and methods, there is no explicit statement about releasing code, nor any repository links or indications that code is in supplementary materials for the methodology described.
Open Datasets Yes Comprehensive experiments conducted on the MOT17 [Milan et al., 2016] and MOT20 [Dendorfer et al., 2020] datasets demonstrate that our approach significantly outperforms SOTA attacks up to 11.8 times, revealing the vulnerability of the tracking-by-detection framework to Bank Tweak.
Dataset Splits Yes Each dataset is split into two halves: one for training the considered detection model and the other for evaluation. The MOT17 and MOT20 datasets are further divided into 30-frame segments, yielding 83 and 148 segments, respectively. Experiments target each segment s (15 19)-th frames for attacks to accumulate features in the objects feature banks over five frames, ensuring accurate evaluation of Bank Tweak s potential effects in practical tracking applications.
Hardware Specification No The paper mentions various software components and models like YOLOX, Faster R-CNN, DETR, OSNet, Res Net, Mobile Net, MLFN, Deep SORT, Strong SORT, and MOTDT, but does not provide any specific details about the hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper lists several software components and models (YOLOX, Faster R-CNN, DETR, OSNet, Res Net, Mobile Net, MLFN, Deep SORT, Strong SORT, MOTDT) but does not provide specific version numbers for any of them. It also mentions attack parameters and iteration counts, which are not software dependencies.
Experiment Setup Yes The feature-based matching threshold is λapp = 0.2, and Io U-based matching threshold is λIo U = 0.7. Attack parameters are ϵ = 4/255 and α = 1/255. In Bank Tweak, dissimilarity loss Ld succeeds when feature similarity exceeds λapp = 0.2, and similarity loss Ls requires cosine distance to be less than λapp = 0.2. Empirically, iterations for Ld are set to Rd = 10, and for Ls, Rs = 150.