Jiantao Zhou

University of Macau
4 previous affiliations
298Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Jiantao Zhou focuses on multimedia forensics and security, multimedia signal processing, and artificial intelligence. His trajectory spans early work in cryptographic primitives and image compression to modern deep learning architectures for image restoration and adversarial robustness. Most recently, his research addresses the security of foundation models, machine unlearning, and advanced video processing techniques.

Research Trajectory

Zhou's early research investigated the security and efficiency of multimedia encryption and compression. In 2006, he analyzed the vulnerabilities of multiple Huffman table-based encryption, proposing chosen-plaintext attacks to highlight the need for careful key selection in On the Security of Multimedia Encryption Schemes Based on Multiple Huffman Table (MHT). This interest in secure and efficient signal processing evolved into the design of joint encryption-and-compression systems. He developed a soft-decision estimation approach for the l2 Restoration of l∞-Decoded Images Via Soft-Decision Estimation, contributing to the broader area of Quantization Interval Constraints for Signal Reconstruction. Building on this foundation, he designed a highly efficient system that allowed encrypted images to be compressed using prediction error clustering in Designing an Efficient Image Encryption-Then-Compression System via Prediction Error Clustering and Random Permutation. His work during this period also explored the intersection of sampling and security, such as his group's development of Bi-level Protected Compressive Sampling, which aligns with the research direction of Simultaneous Compression and Encryption via Compressive Sensing.

As deep learning transformed computer vision, Zhou's focus expanded to advanced image restoration techniques. He introduced a novel rank residual constraint model in From Rank Estimation to Rank Approximation: Rank Residual Constraint for Image Restoration, which progressively approximated low-rank matrices to preserve visual details, advancing the area of Automated Regularization Tuning for Inverse Problems. Recognizing the limitations of early vision transformers in utilizing spatial information, he and his collaborators proposed the Hybrid Attention Transformer (HAT) in Activating More Pixels in Image Super-Resolution Transformer. This architecture combined channel attention with window-based self-attention, contributing significantly to work in Overlapping Window Attention for Spatially Continuous Vision Transformers. The HAT framework was later formalized and expanded in HAT: Hybrid Attention Transformer for Image Restoration, further establishing his contributions to the research direction of Spatial and Channel Attention for Robust Feature Modulation.

Concurrently, Zhou maintained a strong focus on multimedia forensics and the security of machine learning models. He addressed the intellectual property protection of deep neural networks by demonstrating vulnerabilities in passport-based defenses through Effective Ambiguity Attack Against Passport-based DNN Intellectual Property Protection Schemes through Fully Connected Layer Substitution. In the domain of social media forensics, his team proposed a framework for Social Network Analytic-Based Online Counterfeit Seller Detection using User Shared Images, leveraging network measurements like centrality to identify malicious actors when direct product inspection is impossible. His research into adversarial machine learning yielded novel defense and attack strategies. To improve model robustness, his group developed Dual Adversarial Training in DAT: Improving Adversarial Robustness via Generative Amplitude Mix-up in Frequency Domain, which forces models to rely on robust semantic features by mixing stylistic amplitudes. Conversely, he explored the efficacy of black-box attacks by disentangling adversarial and visual features in DifAttack: Query-Efficient Black-Box Adversarial Attack via Disentangled Feature Space, drastically reducing the number of queries required for successful perturbations.

In his most recent work, Zhou has tackled complex, physically-grounded restoration challenges and the security of generative models. He addressed the degradation caused by under-display cameras by explicitly modeling the display as a homogeneous scattering medium in Under-display camera image restoration with scattering effect. His work on moiré pattern removal spans both static images and video; he proposed jointly utilizing RAW and sRGB data in Image Demoiréing in RAW and sRGB Domains to correct color casts, and introduced the DTNet framework in Direction-Aware Video Demoiréing with Temporal-Guided Bilateral Learning, advancing the area of Robust Temporal Alignment for Video Restoration. In the realm of generative AI, his group developed FontGuard: A Robust Font Watermarking Approach Leveraging Deep Font Knowledge to embed information directly into visual font styles. Furthermore, he investigated the mechanics of machine unlearning in large language models, identifying the "squeezing effect" where suppressed responses shift to semantic rephrasings, and proposed a bootstrapping framework to ensure thorough forgetting in LLM Unlearning with LLM Beliefs.

Current Directions

Recent work investigates the intersection of physical image formation models and deep restoration networks, particularly focusing on complex degradations like moiré patterns and scattering effects from under-display cameras. By integrating domain-specific knowledge—such as RAW sensor data and temporal bilateral learning—this research aims to recover high-fidelity visual information in challenging capture scenarios.

Parallel to his restoration work, Zhou's current focus extends to the security, privacy, and intellectual property protection of foundation models. The most recent papers in the evidence address the robust watermarking of generative content and the precise unlearning of specific knowledge in large language models. This involves developing techniques that prevent adversarial evasion, such as mitigating the "squeezing effect" during model unlearning, ensuring that large-scale AI systems remain secure, attributable, and compliant with privacy constraints.

Profile updated

See something wrong? Report a mistake