| MS Seminar

Name of the Speaker: Ms. Janani S (EE22S079)
Guide: Dr. Sheetal Kalyani
Online meeting link: http://meet.google.com/ikk-aapv-qpt
Date/Time: 20th May 2025 (Tuesday), 3:00 PM
Title: Toward Resilient and Efficient Deep Learning: Adversarial Mitigation and Adaptive Scheduling.

Abstract :

The first problem centers on improving adversarial robustness without relying on computationally expensive adversarial training (AT). We propose a novel architectural component—the Adversarial Noise Filter (ANF)—which, when used as the first layer of a network, acts as an implicit defense mechanism against adversarial perturbations. This layer incorporates a large convolutional kernel, an increased number of filters, and a max-pooling operation to suppress high-frequency noise. When integrated into standard models like ResNet, VGG, and EfficientNet, the ANF significantly enhances adversarial robustness. Our method outperforms existing naturally robust architectures and achieves performance competitive with AT-based models across various datasets. Our experimental analysis further demonstrates that ANF-equipped models exhibit (a) wider decision margins, (b) smoother loss surfaces, (c) higher modified peak signal-to-noise ratios (mPSNR) at the ANF output, (d) stronger attenuation of high-frequency adversarial components, and (e) improved denoising under Gaussian noise compared to baseline networks. In the second part of the talk, we introduce the Probabilistic Learning Rate Scheduler (PLRS)—a non-monotonic, stochastic scheduler designed with formal convergence guarantees. PLRS allows for probabilistic oscillations during training in contrast to learning rate schedulers such as cosine annealing. Experiments across a variety of neural architectures and datasets show that PLRS performs on par with or better than state-of-the-art schedulers