Deep Learning Under Siege: Identifying Security Vulnerabilities and Risk Mitigation Strategies

TL;DR

The paper addresses security vulnerabilities in deep learning architectures deployed in production and anticipates future risks as computing, AI, and hardware evolve. It develops a taxonomy of attacks with a structured naming convention ($\alpha,\beta,\ldots$) and proposes risk mitigation policies to reduce exposure. It also defines metrics to quantify the effectiveness of mitigations—such as adversarial robustness $R$, data quality measures $DQ_{filtered}$, anomaly/monitoring rates, and privacy loss $P$ bounded by a threshold $\mathcal{E}$—and demonstrates example calculations. The work aims to aid policymakers, risk managers, and practitioners in planning secure DL deployments and suggests future work on generating data to train DL models to detect DL security threats.

Abstract

With the rise in the wholesale adoption of Deep Learning (DL) models in nearly all aspects of society, a unique set of challenges is imposed. Primarily centered around the architectures of these models, these risks pose a significant challenge, and addressing these challenges is key to their successful implementation and usage in the future. In this research, we present the security challenges associated with the current DL models deployed into production, as well as anticipate the challenges of future DL technologies based on the advancements in computing, AI, and hardware technologies. In addition, we propose risk mitigation techniques to inhibit these challenges and provide metrical evaluations to measure the effectiveness of these metrics.