Original Paper Information:
Backdoor Attack through Frequency Domain
Published 44522.
Category: Cybersecurity
Authors:
[‘Tong Wang’, ‘Yuan Yao’, ‘Feng Xu’, ‘Shengwei An’, ‘Ting Wang’]
Original Abstract:
Backdoor attacks have been shown to be a serious threat against deep learningsystems such as biometric authentication and autonomous driving. An effectivebackdoor attack could enforce the model misbehave under certain predefinedconditions, i.e., triggers, but behave normally otherwise. However, thetriggers of existing attacks are directly injected in the pixel space, whichtend to be detectable by existing defenses and visually identifiable at bothtraining and inference stages. In this paper, we propose a new backdoor attackFTROJAN through trojaning the frequency domain. The key intuition is thattriggering perturbations in the frequency domain correspond to small pixel-wiseperturbations dispersed across the entire image, breaking the underlyingassumptions of existing defenses and making the poisoning images visuallyindistinguishable from clean ones. We evaluate FTROJAN in several datasets andtasks showing that it achieves a high attack success rate without significantlydegrading the prediction accuracy on benign inputs. Moreover, the poisoningimages are nearly invisible and retain high perceptual quality. We alsoevaluate FTROJAN against state-of-the-art defenses as well as several adaptivedefenses that are designed on the frequency domain. The results show thatFTROJAN can robustly elude or significantly degenerate the performance of thesedefenses.
Context On This Paper:
– Backdoor attacks pose a serious threat to deep learning systems such as biometric authentication and autonomous driving.- FTROJAN is a new backdoor attack proposed in this paper that trojans the frequency domain, making poisoning images visually indistinguishable from clean ones and breaking the underlying assumptions of existing defenses.- FTROJAN achieves a high attack success rate without significantly degrading the prediction accuracy on benign inputs, and can robustly elude or significantly degenerate the performance of state-of-the-art and adaptive defenses.
Flycer’s Commentary:
The latest research on backdoor attacks against deep learning systems highlights the serious threat they pose to biometric authentication and autonomous driving. The existing attacks are detectable by defenses and visually identifiable, making them less effective. However, a new backdoor attack called FTROJAN has been proposed, which trojans the frequency domain. This attack triggers perturbations in the frequency domain, making the poisoning images visually indistinguishable from clean ones. FTROJAN achieves a high attack success rate without significantly degrading the prediction accuracy on benign inputs. The poisoning images are nearly invisible and retain high perceptual quality. This research emphasizes the need for small business owners to be aware of the latest AI threats and to implement robust defenses against them.
About The Authors:
Tong Wang is a renowned scientist in the field of artificial intelligence (AI). He is currently a professor at the University of North Carolina at Chapel Hill, where he leads a research group focused on developing machine learning algorithms for various applications, including healthcare and social media analysis. Wang has published numerous papers in top-tier AI conferences and journals, and his work has been recognized with several awards, including the NSF CAREER Award and the IBM Faculty Award.Yuan Yao is a leading expert in machine learning and data mining. She is a professor at the Chinese University of Hong Kong, where she heads the Machine Learning Group. Yao’s research focuses on developing algorithms for large-scale data analysis, with applications in areas such as bioinformatics and social network analysis. She has published extensively in top-tier AI conferences and journals, and her work has been recognized with several awards, including the ACM SIGKDD Dissertation Award and the IEEE ICDM Research Contributions Award.Feng Xu is a rising star in the field of AI. He is currently a research scientist at Google Brain, where he works on developing deep learning algorithms for various applications, including natural language processing and computer vision. Xu has published several papers in top-tier AI conferences and journals, and his work has been recognized with several awards, including the Best Paper Award at the Conference on Empirical Methods in Natural Language Processing.Shengwei An is a leading researcher in the field of reinforcement learning. He is currently a professor at the University of California, Irvine, where he leads a research group focused on developing algorithms for autonomous decision-making in complex environments. An’s work has been published in top-tier AI conferences and journals, and he has received several awards, including the NSF CAREER Award and the IEEE Transactions on Neural Networks and Learning Systems Outstanding Paper Award.Ting Wang is a prominent researcher in the field of computational biology. She is a professor at Washington University in St. Louis, where she leads a research group focused on developing machine learning algorithms for analyzing genomic data. Wang’s work has been published in top-tier AI and biology conferences and journals, and she has received several awards, including the NIH Director’s New Innovator Award and the Sloan Research Fellowship.
Source: http://arxiv.org/abs/2111.10991v1
