Explore advanced computer vision concepts through this comprehensive graduate-level course from the University of Central Florida covering cutting-edge research in adversarial machine learning, explainable AI, and neural network robustness. Delve into adversarial attacks and defenses, examining topics such as motion-excited samplers for video adversarial attacks, adversarial contrastive learning methods, and cross-domain transferability of adversarial perturbations. Master explainable AI techniques including SmoothGrad for noise reduction, meaningful perturbation methods for interpretable explanations, and empirical studies of deep neural network explanation methods. Investigate neural network robustness through analysis of universal adversarial perturbations, feature denoising approaches, and adaptive attacks on adversarial example defenses. Study foundational papers on intriguing properties of neural networks, transferability improvements through input diversity, and the fundamental principles behind explaining and harnessing adversarial examples. Gain expertise in evaluating neural network robustness, understanding skip connection impacts on transferable adversarial examples, and developing models resistant to adversarial attacks through this research-focused curriculum spanning over 12 hours of advanced computer vision content.

Syllabus

CAP6412 21Spring-Introduction Lecture -1
CAP6412 21Spring-Introduction Lecture -2
CAP6412 21Spring- Motion-Excited Sampler: Video Adversarial Attack with Sparked Prior
CAP6412 21Spring-Robust Pre-Training by Adversarial Contrastive Learning
CAP6412 21Spring-Contrastive Learning with Adversarial Examples
CAP6412 21Spring-Adversarial Self-Supervised Contrastive Learning
CAP6412 21Spring-How Can I Explain This to You? An Emp. Study of Deep Neural Net Explanation Methods
CAP6412 21Spring-Explainable AI: Interpreting, Explaining and Visualizing Deep Learning
CAP6412 21Spring-Smoothgrad: removing noise by adding noise
CAP6412 21Spring-Unmasking clever Hans predictors and assessing what machines really learn
CAP6412 21Spring-Interpretable explanations of black boxes by meaningful perturbation
CAP6412 21Spring-Cross-domain transferability of adversarial perturbations
CAP6412 21Spring-Fast is better than free: Revisiting adversarial training
CAP6412 21Spring-Evading defenses to transferable ad. examples by translation-invariant attacks
CAP6412 21Spring-Feature denoising for improving adversarial robustness
CAP6412 21Spring-Skip connections matter: On the transferability of ad. ex. generated with resnets
CAP6412 21Spring- On adaptive attacks to adversarial example defenses
CAP6412 21Spring-Towards deep learning models resistant to adversarial attacks
CAP6412 21Spring-Universal adversarial perturbations
CAP6412 21Spring-Towards Evaluating the Robustness of Neural Networks
CAP6412 21Spring-Improving transferability of adversarial examples with input diversity
CAP6412 21Spring-Explaining and harnessing adversarial examples
CAP6412 21Spring-Intriguing properties of neural networks