作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• •    

基于双域扰动响应的图像后门样本检测方法

  • 发布日期:2026-11-18

Image Backdoor Sample Detection Based on Dual-Domain Perturbation Responses

  • Published:2026-11-18

摘要: 针对不可见后门触发器可能分布于图像空间模式或低位比特编码、单一扰动探针难以在不同攻击机制间稳定分离干净样本与投毒样本的问题,提出一种基于双域扰动响应的图像后门样本检测方法(PB-Fusion)。该方法面向污染训练样本筛查,在不修改待检测模型参数且不依赖触发器图案、位置和目标标签先验的条件下,从像素域与比特平面域构造互补探针。像素域采用待测图像与随机干净参考图像的等权混合,以观察宏观空间与纹理结构受扰后的预测变化;比特平面域将参考图像指定低位翻转后嵌入待测图像,并对嵌入差分进行缩放,以激励模型对细粒度量化编码变化的响应。针对每个样本在两域各执行100次随机探测,分别计算平均预测分布的预测熵以及预测熵与单次熵均值之差所定义的互信息,形成每域二维、共四维的响应特征。随后,仅利用干净拟合集计算各维特征的中位数与中位绝对偏差,构造稳健域异常分数,并按0.6和0.4的权重融合像素域与比特平面域分数;检测阈值由独立干净校准集在目标假阳性率0.10处确定。实验在含10%投毒样本的CIFAR-10完整50000张训练池和Tiny-ImageNet完整100000张训练池上进行,覆盖BadNets、Blended、FTrojan、SIG和LBP-LSB五类攻击,并对IBD-PSC、PSBD和PB-Fusion采用相同受害模型与样本进行3次独立重复实验。结果表明,PB-Fusion在全部数据集与攻击组合上的宏平均受试者工作特征曲线下面积(AUC)、真阳性率(TPR)和假阳性率(FPR)分别为0.9826±0.0020、0.9728±0.0010和0.1119±0.0014;其总体TPR较IBD-PSC和PSBD分别高16.29和17.21个百分点,总体AUC分别高0.1124和0.0415。在CIFAR-10和Tiny-ImageNet上的平均TPR分别达到0.9611±0.0090和0.9845±0.0109。特征域消融显示,双域融合在10个数据集-攻击组合中的6个组合取得最高AUC、9个组合与最佳单域AUC差距不超过0.01,并在8个组合取得最高TPR;融合权重敏感性分析进一步表明,预先固定的0.6像素域权重处于较优且稳定的性能区间。上述结果说明,双域响应的稳健加权融合能够获得较高的跨攻击检出率,并将总体FPR维持在约0.11。

Abstract: Invisible backdoor triggers may be encoded either in image-space patterns or in low-bit representations, making a single probing domain unreliable across attacks. To address this problem, an image backdoor sample detection method based on dual-domain perturbation responses (PB-Fusion) is proposed. Without modifying the inspected model or requiring prior knowledge of the trigger pattern, location, or target label, PB-Fusion constructs complementary probes in the pixel and bit-plane domains. The pixel probe equally mixes a candidate image with a randomly sampled clean reference image to expose changes in spatial and texture responses. The bit-plane probe flips selected low bits of a clean reference image, embeds them into the candidate image, and scales the embedding difference to stimulate responses to fine-grained quantization patterns. For each candidate, 100 random probes are evaluated in each domain. Predictive entropy and mutual information form a two-dimensional response feature for each domain. Robust domain scores are computed from the median and median absolute deviation estimated on a clean fitting set, fused with pixel and bit-plane weights of 0.6 and 0.4, and calibrated on an independent clean set at a target false positive rate of 0.10. Experiments screen the complete CIFAR-10 training pool of 50,000 images and the complete Tiny-ImageNet training pool of 100,000 images, each containing 10% poisoned samples, under BadNets, Blended, FTrojan, SIG, and LBP-LSB.IBD-PSC, PSBD, and PB-Fusion are evaluated on the same victim models and samples over three independent runs. Across all dataset–attack combinations, the macro-average values of the area under the receiver operating characteristic curve (AUC), true positive rate (TPR), and false positive rate (FPR) for PB-Fusion are 0.9826±0.0020, 0.9728±0.0010, and 0.1119±0.0014, respectively. Its overall TPR exceeds those of IBD-PSC and PSBD by 16.29 and 17.21 percentage points, while its overall AUC is higher by 0.1124 and 0.0415. The average TPR values on CIFAR-10 and Tiny-ImageNet reach 0.9611±0.0090 and 0.9845±0.0109. Domain ablation shows that fusion obtains the highest AUC in 6 of 10 dataset-attack combinations and remains within 0.01 of the best single-domain AUC in 9 combinations; it also achieves the highest TPR in 8 combinations. Fusion-weight analysis confirms stable performance around the fixed pixel-domain weight of 0.6. These results demonstrate high cross-attack detection rates with an overall FPR of approximately 0.11.