From Head to Tail: Efficient Black-box Model Inversion Attack via Long-tailed Learning

Ziang Li, Hongguang Zhang, Juan Wang*, Meihui Chen, Hongxin Hu, Wenzhe Yi, Xiaoyang Xu, Mengda Yang, Chenjun Ma

• We introduce SMILE, an efficient black-box MIA. By combining long-tailed surrogate training and gradient-free black-box optimization, SMILE outperforms existing black-box MIAs with about 5% of the query overhead.

CCall: Recovering Indirect Call Targets from Binaries With Cross-Domain Fine-Tuning

Bin Weng, Yunru Wang, juan Wang*, Mengda Yang, Ziang Li, Fei Li

• We propose a novel cross-domain fine-tuning strategy based on domain adaptation, which can further study the semantics from the unlabeled test samples. This cross-domain finetuning strategy can also be applied in other AI-based downstream binary analysis tasks.

Penetralium: Privacy-preserving and memory-efficient neural network inference at the edge

Mengda Yang, Wenzhe Yi, Juan Wang*, Hongxin Hu, Xiaoyang Xu, Ziang Li

• Penetralium is a novel model inference system that we design to provide robust security for deep learning computation at the edge. Penetralium is created with system and algorithm co-design, and has little overhead and no impact on prediction accuracy.

A Stealthy Wrongdoer: Feature-Oriented Reconstruction Attack against Split Learning

Xiaoyang Xu, Mengda Yang, Wenzhe Yi, Ziang Li, Juan Wang, Hongxin Hu, Yong Zhuang, Yaxin Liu

• We propose a novel attack, named Feature-Oriented Reconstruction Attack (FORA). As far as we know, FORA is the first work enabling a semi-honest server to perform powerful DRA in more realistic and challenging SL systems.

GAN You See Me? Enhanced Data Reconstruction Attacks against Split Inference

Ziang Li, Mengda Yang, Yaxin Liu, Juan Wang*, Hongxin Hu, Wenzhe Yi, Xiaoyang Xu

• We propose GLASS and GLASS++, which are enhanced DRAs combined with pre-trained StyleGAN models. We conduct a systematic evaluation and comparison of various DRAs against seven defense mechanisms.

Measuring Data Reconstruction Defenses in Collaborative Inference Systems

Mengda Yang, Ziang Li, Juan Wang*, Hongxin Hu, Ao Ren, Xiaoyang Xu, Wenzhe Yi

• We are the first to experimentally verify the robustness of reconstruction defenses for inference data privacy in collaborative systems. We devise a technique called SFD against the existing defense mechanisms.

ProcGuard: Process Injection Behaviours Detection Using Fine-grained Analysis of API Call Chain with Deep Learning

Juan Wang*, Chenjun Ma, Ziang Li, Huanyu Yuan, Jie Wang

• We present a framework for detecting process injection attacks called ProcGuard, which adopts API call chain analysis and deep learning.