The Trust Tax of Privacy and Robustness: An Empirical Study Across Vision, NLP, and Tabular ML
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
While adversarial robustness and differential privacy are recognized as vital for trustworthy machine learning, the systems-level costs of these features remain unsystematized and poorly understood. The performance-centric paradigm of ML systems, exemplified by benchmarks like MLPerf, has been structurally blind to the unique computational patterns of trustworthy workloads, creating a significant knowledge gap for practitioners and hardware designers. This paper presents the cross-domain, energy-aware measurement study of the hidden“trust tax”in deep learning. We evaluate three representative tasks: vision (ResNet-18), NLP (DistilBERT), and tabular (MLP). We quantify the cost of standard privacy (DP-SGD) and robustness (PGD) defenses. On a single NVIDIA V100 GPU, we find this tax is steep: PGD adversarial training increases wall-time and energy by$4.07 \times$, while DP-SGD$(\epsilon=8)$raises the cost by$3.55 \times$and slashes clean accuracy from$\text{8 7 \%}$to$\approx \text{56\%}$. Our micro-architectural profiling reveals the root cause of this tax: trust algorithms like per-example gradient clipping and iterative attacks under-utilize specialized hardware like tensor cores, creating memory-bound bottlenecks. By providing the cross-domain systematization of these costs, our work serves as a foundational reference and public dataset, laying the empirical groundwork for the next generation of trust-aware compilers, schedulers, and hardware.
Fields of Study & Subjects
Adversarial Robustness in Machine Learning
Security and Verification in Computing
Privacy-Preserving Technologies in Data
Publication details
- Journal / source
- International Conference on Artificial Intelligence in Information and Communication (ICAIIC)
- Pages
- 992–996
- Publication date
- 24 February 2026
- Language
- EN