This talk gives a brief review of Wang group’s current research efforts at UBC, in the areas of Adversarial Deep Learning in Digital Media Security & Forensics. Deep learning has achieved state-of-the-art performances in many applications. Unfortunately, current deep learning models however could be sensitive to perturbations, giving rise to security, privacy and reliability issues in practical applications.

Under the paradigm of adversarial deep learning, as an attacker, we study potential adversarial attacks and explore novel approaches to scrutinize potential vulnerabilities of deep learning models in digital media security & forensics, by investigating three fundamental learning tasks: matching, classification and regression. Specifically, this talk presents novel attacks (both in the digital domain and in the physical domain) for several essential models belonging to the above three dominant tasks: 1) image hashing for image retrieval and authentication, as a typical matching task; 2) GAN-generated fake face imagery forensics, as a representative binary classification task; 3) multiclass image classification; 4) camera-LIDAR 3d object detection; and 5) single object tracking in videos, which is an important video surveillance model involving a combination of the matching task, the classification task and the regression task. We address security and privacy threats that arise in the above typical digital media problems and study how to fool deep learning models to make wrong decisions.