Building Safe & Resource Efficient NN through architectural collapse
Year: 2026
Type: article
Abstract: Modern Convolutional Neural Networks (CNNs) often suffer from depth redundancy, where sequential layers increase inference latency and memory usage without significantly improving accuracy. To address this, we introduce Layer-by-Layer Structural Reduction (LBL), a post-hoc methodology that identifies contiguous layer blocks and collapses them into a single, lightweight surrogate module. By targeting depth redundancy directly, LBL aims to reduce sequential operations and increase hardware efficiency. While LBL streamlines networks for deployment, ensuring these resource-efficient models remain secure is critical. We are working on evaluating model resilience using Adversarial Particle Swarm Optimization (APSO), a data-agnostic, black-box framework that generates transferable adversarial examples to probe cross-model vulnerabilities. This ongoing work investigates how combining structural collapse with reproducible robustness metrics and Explainable AI (XAI) diagnostics can establish a unified paradigm for building and validating deep learning systems that are both computationally efficient and resilient to inference-time poisoning.
Author: Gafur, Jamil
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7128364183
FWCI:
Cites:
Cited by:
Related to: