Projects

1. Transferable Adversarial Attacks for Image and Video Object Detection

In this project, we aims at proposing an Unified and Efficient Adversary (UEA) for object detection. UEA can simultaneously attack two kinds of representative object detectors: proposal based detectors like Faster-rcnn, and regression based detector like SSD. In addition, UEA use a generative mechanism to generate adversarial examples, therefore, can efficiently perturbs all the frames in a video. Thus, UEA can attack image and video object detection.

2. Sparse Adversarial Perturbations for Videos

In this project, we try to attack the action recognition task. Towards this end, we choose a classic method for action recognition: CNN+LSTM. The proposed method is an optimization process based on the L2,1 norm. Under the L2,1 constriant, adversarial perturbations are added on the sparse frames. Moreover, the adversarial perturbations will propgate to the clean frames with the help of LSTM. In this way, the sparse perturbations will pollute all the frames in a video, and resulting in the wrong predicted label for action recognition.