Developing novel deep learning architectures that enable image authenticity detection while preserving the privacy of individuals. This project combines differential privacy techniques with state-of-the-art forensic detection methods to create privacy-conscious verification systems.

Key Contributions:

• Novel privacy-preserving CNN architecture with formal privacy guarantees
• Benchmark dataset for private forensics evaluation
• 15% improvement in privacy-utility tradeoff compared to baselines
• Published at CVPR 2024 (Best Paper Honorable Mention)

Investigating federated learning approaches for distributed image authentication across multiple organizations without sharing sensitive data. This work addresses the challenge of building robust models from decentralized data sources while maintaining data privacy.

Key Contributions:

• Federated aggregation strategy specialized for forensic features
• Communication-efficient training protocol reducing bandwidth by 60%
• Multi-institutional validation framework with 5 partner organizations
• Under review at IEEE TIFS

Examining the vulnerabilities of deepfake detection systems to adversarial attacks and developing robust defense mechanisms. This project explores both white-box and black-box attack scenarios to strengthen detection reliability in real-world applications.

Key Contributions:

• Comprehensive adversarial benchmark with 10+ attack methods
• Robust training methodology improving adversarial accuracy by 25%
• Certified defense mechanism with provable guarantees
• Published at ECCV 2023 and IEEE TPAMI 2023

Leveraging multiple modalities (visual, audio, metadata) for comprehensive media authenticity verification. This interdisciplinary project combines computer vision with signal processing to detect sophisticated multimedia manipulations.

Key Contributions:

• Cross-modal fusion architecture for synchronized analysis
• Large-scale multi-modal dataset with 50K+ samples
• Attention-based integration mechanism achieving 92% detection accuracy
• Published at ICML 2023

Creating high-quality benchmark datasets to facilitate research in privacy-preserving and robust computer vision. These datasets include diverse manipulation types, privacy annotations, and evaluation protocols.

Key Contributions:

• Privacy-annotated forensics dataset with 100K+ images
• Standardized evaluation protocol adopted by 15+ research groups
• Diverse manipulation types covering real-world scenarios
• Published at NeurIPS 2022 (Datasets Track)

Developing efficient architectures for real-time deepfake detection suitable for deployment in resource-constrained environments. Focus on model compression, edge computing, and efficient inference strategies.

Key Contributions:

• Lightweight detection model with 10x speedup over baselines
• Edge-optimized architecture for mobile deployment
• Knowledge distillation framework maintaining 95% accuracy
• Work in progress - submission planned for CVPR 2025