Past Recording
ShareSave
Paper Review - Threat of Adversarial Attacks on Deep Learning in Computer Vision: A Survey
Wednesday Jun 24 2020 14:00 GMT
Please to join the live chat.
Paper Review - Threat of Adversarial Attacks on Deep Learning in Computer Vision: A Survey
Why This Is Interesting

With increased dependence on computer vision algorithms to support autonomous driving, it is important to understand the vulnerabilities and threats associated with these algorithms; their impact on safety and security. Dr. Saini will present a conceptual review based on the survey paper: Threat of Adversarial Attacks on Deep Learning in Computer Vision.

Discussion Points

1)Can you expand on where do you see the near-term need for additional research in the area of computer vision to support autonomous driving? 2) Do we have any benchmark datasets that have been created specifically for the domain of autonomous driving and could be used to validate robustness to attacks? 3)Can you talk a little bit about the privacy implications for using computer vision algorithms to support autonomous driving?

Takeaways

Despite the high accuracies of deep neural networks on a wide variety of Computer Vision tasks, these are vulnerable to subtle input perturbations that lead them to completely change their outputs

It is apparent that adversarial attacks are a real threat to deep learning in practice, especially in safety and security-critical applications

The existing literature demonstrates that currently deep learning can be effectively attacked in cyberspace as well as in the physical world

Time of Recording: Wednesday Jun 24 2020 14:00 GMT
slides: please to see content