Bi-Weekly AI Research Roundup

Latest research summaries in ML, Robotics, CV, NLP and AI

Contents

1. ChatGPT's Potential in Cryptography Misuse Detection: A Comparative Analysis with Static Analysis Tools

2. EyeCLIP: A visual-language foundation model for multi-modal ophthalmic image analysis

3. Generalization of Graph Neural Networks is Robust to Model Mismatch

4. LLaMA-Omni: Seamless Speech Interaction with Large Language Models

5. Human Perception of LLM-generated Text Content in Social Media Environments

6. Demo: SGCode: A Flexible Prompt-Optimizing System for Secure Generation of Code

7. VLMEvalKit: An Open-Source Toolkit for Evaluating Large Multi-Modality Models

8. Approximation and generalization properties of the random projection classification method

9. Synthetic continued pretraining

10. Awaking the Slides: A Tuning-free and Knowledge-regulated AI Tutoring System via Language Model Coordination

11. AI-accelerated discovery of high critical temperature superconductors

12. DreamHOI: Subject-Driven Generation of 3D Human-Object Interactions with Diffusion Priors

13. Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-Based Decoding

14. Source2Synth: Synthetic Data Generation and Curation Grounded in Real Data Sources

15. OmniQuery: Contextually Augmenting Captured Multimodal Memory to Enable Personal Question Answering

---

ChatGPT's Potential in Cryptography Misuse Detection: A Comparative Analysis with Static Analysis Tools

Authors: Ehsan Firouzi, Mohammad Ghafari, Mike Ebrahimi

Source and references: https://arxiv.org/abs/2409.06561v1

---

Introduction

This paper investigates the potential of ChatGPT, a large language model, in detecting cryptography misuses in the Java Cryptography Architecture (JCA).

Key Points

• The paper evaluates ChatGPT's performance in detecting a variety of cryptography misuses and compares it to the state-of-the-art static analysis tool, CryptoGuard.

• The authors use the CryptoAPI-Bench benchmark to assess ChatGPT's capabilities and apply prompt engineering techniques to enhance its detection abilities.

• The authors also evaluate the generalizability of their findings using a different benchmark, CAMBench.

Methodology

The researchers followed a four-step process to assess ChatGPT's JCA misuse detection capabilities. This included compiling security violation rules, analyzing gaps in the CryptoAPI-Bench, evaluating test cases with ChatGPT, and assessing additional code snippets. To improve ChatGPT's performance, the authors applied prompt engineering techniques based on insights from recent research.

Results and Findings

The initial results showed that ChatGPT achieved an average F-measure of 86% across 12 misuse categories, outperforming CryptoGuard in 5 categories. After prompt engineering, ChatGPT's average F-measure improved to 94.6%, outperforming CryptoGuard in 10 categories. The authors also confirmed the generalizability of their optimized prompts using the CAMBench.

Implications and Conclusions

The study highlights ChatGPT's promising potential for detecting cryptography misuse, with the ability to outperform the state-of-the-art static analysis tool, CryptoGuard, through the use of prompt engineering. This suggests that large language models like ChatGPT can be leveraged to enhance security practices in software development.

---

EyeCLIP: A visual-language foundation model for multi-modal ophthalmic image analysis

Authors: Danli Shi, Weiyi Zhang, Jiancheng Yang, Siyu Huang, Xiaolan Chen, Mayinuer Yusufu, Kai Jin, Shan Lin, Shunming Liu, Qing Zhang, Mingguang He

Source and references: https://arxiv.org/abs/2409.06644v1

---

Introduction

This research paper presents EyeCLIP, a visual-language foundation model developed for multi-modal ophthalmic image analysis. EyeCLIP aims to leverage real-world multi-examination data and language information to enhance the analysis and diagnosis of ophthalmic and systemic diseases.