Year: 2025
Type: article
Abstract: Deep neural networks (DNNs) excel in image classification but are vulnerable to backdoor attacks due to reliance on external training data, where specific markers trigger preset misclassifications. Existing attack techniques have an obvious trade-off between the effectiveness of the triggers and the stealthiness, which limits their practical application. For this purpose, in this paper, we develop a method—chromatic channel-based implicit backdoor attack (CCIBA), which combines a discrete wavelet transform (DWT) and singular value decomposition (SVD) to embed triggers in the frequency domain through the chromaticity properties of the YUV color space. Experimental validation on different image datasets shows that compared to existing methods, CCIBA can achieve a higher attack success rate without a large impact on the normal classification ability of the model, and its good stealthiness is verified by manual detection as well as different experimental metrics. It successfully circumvents existing defense methods in terms of sustainability. Overall, CCIBA strikes a balance between covertness, effectiveness, robustness and sustainability.
Source: Electronics
Authors: Chaoliang Li, Jiyan Liu, Yang Liu, Shengjie Yang
Institution: Hunan University of Technology and Business
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W4414145894