Summary of Morel: Enhancing Adversarial Robustness Through Multi-objective Representation Learning, by Sedjro Salomon Hotegni et al.
MOREL: Enhancing Adversarial Robustness through Multi-Objective Representation Learning
by Sedjro Salomon Hotegni, Sebastian Peitz
First submitted to arxiv on: 2 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computer Vision and Pattern Recognition (cs.CV)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed MOREL approach enhances the robustness of deep neural networks (DNNs) against white-box and black-box adversarial attacks by learning strong feature representations during training. The method uses cosine similarity loss and multi-positive contrastive loss to align natural and adversarial features from the model encoder, ensuring tight clustering. This is achieved through an embedding space where the classifier is motivated to achieve accurate predictions while the DNN is trained to produce similar features for inputs within the same class, despite perturbations. The approach outperforms other methods that require no architectural changes or test-time data purification. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary MOREL helps deep neural networks (DNNs) be more resilient against tiny changes in input data that can trick them into making wrong predictions. This is a big problem because it means DNNs are not as reliable as we thought. To solve this, researchers developed a way to train DNNs so they learn features that are robust against these small changes. The new approach uses two types of losses: one that makes the model’s embedding space look like a tight cluster and another that helps the classifier make accurate predictions. This results in DNNs that can withstand attacks from hackers trying to trick them. |
Keywords
» Artificial intelligence » Clustering » Contrastive loss » Cosine similarity » Embedding space » Encoder
