Research map / Adversarial robustness and security / Certified and empirical robustness
Towards Trustworthy Deep Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
A AAAI-24 New Faculty Highlights talk surveying the author's work on robust ML (CLEVER, CNN-Cert, SingleProp) and interpretable ML (CLIP-Dissect, Label-free CBM) toward trustworthy deep learning.
Problem
DNNs are vulnerable to adversarial examples and have opaque internal workings, which hinders their trustworthy deployment in mission-critical settings.
Approach
Summarizes a line of work that provides attack-agnostic robustness evaluation and efficient certification for DNNs, and fully automated, label-free neuron interpretation and interpretable concept bottleneck models that keep high accuracy.
Key results
- CLIP-Dissect is reported to be 10-200x faster than existing neuron interpretation methods while giving more accurate descriptions.
- SingleProp is reported to be up to 13.5x faster to train than prior verification-bound-based robust training.
- Label-free CBM is claimed to be the first concept bottleneck model that scales to large-scale ImageNet while retaining high accuracy.
Limitations
- Talk abstract only; no experimental details or full results provided.
- Covers prior works by the author rather than presenting new experiments.
Experimental setup
Builds on: CLIP concept bottleneck model
Closest papers by meaning
- Interpreting Robustness Proofs of Deep Neural Networks (ICLR 2024 · similarity 0.91)
- VNN: Verification-Friendly Neural Networks with Hard Robustness Guarantees (ICML 2024 · similarity 0.89)
- Improving the Adversarial Robustness and Interpretability of Deep Neural Networks by Regularizing Their Input Gradients (AAAI 2018 · similarity 0.88)
- Neural Architecture Design and Robustness: A Dataset (ICLR 2023 · similarity 0.88)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.87)
- SURE: SUrvey REcipes for building reliable and robust deep networks (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.