AI Research Analysis

Exploring LLM Adaptation for CVD

The task is framed as a binary classification problem: identifying if a given source code function is vulnerable (1) or safe (0). Our experiments utilized two datasets:

• BigVul (2020): Derived from real-world GitHub projects, containing 3,754 CVEs and 188,636 C/C++ functions (5.7% vulnerable). Preprocessed for relevant columns, split into training/validation/testing sets, and under-sampled for class balance.
• PrimeVul (2024): A more recent and accurate dataset with 6,968 vulnerable and 228,800 benign functions across 140 CWEs, using its original data split.

We established baselines using state-of-the-art medium-sized models, CodeBERT and UniXcoder, fine-tuning them on our datasets for direct comparison.

LLM Approach: Llama-3.1 8B Base and Instruct Models

Our study on Llama-3.1 8B explored two main categories of adaptation:

1. Prompt Engineering: Adjusting LLM behavior without changing weights.
   • Zero-shot Prompting: Providing only the instruction, relying solely on pre-trained knowledge.
   • Few-shot Prompting: Including 6 labeled examples in the prompt. We tested three example selection strategies: random, same vulnerability type (CWE), and Retrieval Augmented Generation (RAG) using CodeBERT embeddings for similarity.
2. Fine-tuning: Adapting LLMs by training on custom datasets using QLoRA (Quantized Low Rank Adapters) for memory efficiency.
   • Generative Fashion: Fine-tuning the LLM to generate textual responses (e.g., "Vulnerable" or "Safe").
   • Classification Fashion: Adding a feed-forward neural network (FFNN) classification head to the LLM to directly output vulnerability probability.
   • Test-Time Fine-tuning: For each test sample, retrieve 6 similar examples (via RAG) and perform a quick fine-tuning of the model before inference.
   • Double Fine-tuning (Novel): A hybrid approach where the model is first fine-tuned on the entire training dataset (with a classification head), and then further tuned at test-time using RAG-selected examples.

Performance Benchmarks and Key Observations

Our experiments reveal distinct performance differences across various techniques for Llama-3.1 8B.

Baseline Performance

CodeBERT and UniXcoder, when fully fine-tuned on BigVul, achieved excellent F1-scores of 0.92 and 0.94 respectively. On PrimeVul, their performance reduced to 0.74 and 0.77 F1-scores, indicating the more challenging nature of the dataset.

Llama-3.1 Prompting Performance

• Zero-shot Prompting: Achieved a medium 0.514 F1-score for the "vulnerable" class on BigVul, with a bias towards predicting "vulnerable." This suggests general knowledge but a lack of precise detection.
• Few-shot Prompting:
  • Random/Same CWE: Showed some improvement in recognizing "safe" code, but did not significantly boost vulnerable detection.
  • RAG (Retrieval Augmented Generation): Proved most effective among prompting techniques, with a 0.692 F1-score for "vulnerable" class on BigVul. This indicates that relevant examples significantly aid the model.

Llama-3.1 Fine-tuning Performance

• Test-Time Fine-tuning (RAG): Improved detection capacity, yielding a 0.792 F1-score for "vulnerable" class on BigVul, outperforming few-shot RAG. This highlights the benefit of dynamically adapting model weights per test instance.
• Efficient Fine-tuning with QLORA:
  • Generative Fashion (Llama Instruct): Achieved an average F1-score of 0.900 on BigVul.
  • Classification Fashion (Llama Base + Classification Head): Demonstrated superior results, reaching an average F1-score of 0.950 on BigVul, comparable to fine-tuned UniXcoder. This indicates that feeding LLM embeddings into a binary classifier is more effective than text generation for this task.
• Double Fine-tuning: Our novel approach, combining full training with test-time fine-tuning, achieved the best overall performance with a 0.970 F1-score on BigVul, surpassing all baselines. On PrimeVul, it matched the UniXcoder baseline's 0.770 average F1-score.

ROC Curves and t-SNE Plots: Double Fine-tuning achieved an AUC of 0.997 on BigVul, demonstrating near-perfect discrimination. t-SNE plots confirmed that fine-tuning with a classification head enables clear separation of vulnerable and safe code embeddings, which is not observed in untuned or merely prompted models.

Strategic Takeaways & Future Directions

Key Takeaways:

1. Fine-tuning is Crucial: Llama models perform poorly in zero-shot or few-shot prompting for CVD. The task's complexity requires models to gain task-specific knowledge through fine-tuning. RAG is the most effective prompting strategy for example selection, and Test-Time Fine-tuning provides even greater benefits.
2. Llama-3.1 Models Show Strong Potential: Properly fine-tuned Llama 3.1 8B can match or even exceed state-of-the-art CVD models. Its broader pre-training on diverse text, including cybersecurity-related content, suggests a superior capacity for reasoning over complex and subtle vulnerability patterns compared to code-specific models.
3. Versatility Beyond Detection: Llama-3.1's flexibility could extend its utility to broader security workflows, such as vulnerability reasoning and security report generation, consolidating multiple tasks into a single, powerful model.

Future Work Perspectives:

Our research opens several avenues for future exploration:

• Interpretability & Failure Analysis: Investigate model predictions through explainability techniques and conduct per-CWE failure case analysis.
• Cost & Scalability: Analyze the practical deployment concerns of LLMs, including training/inference latency, GPU requirements, carbon footprint, and potential optimizations.
• Advanced Prompting & Diverse Datasets: Experiment with more sophisticated prompting and test on other C/C++ datasets, different programming languages (e.g., memory-safe languages), and emerging Graph LLMs to assess generalization capacity.
• Refining Double Fine-tuning: Further validate the value of our novel approach and distinguish its gains from inherent model capabilities.
• Broader Workflow Integration: Expand research towards multi-class vulnerability classification and integrating LLMs into a complete vulnerability management workflow, including assessment and remediation phases.
• Real-world Exploitability & SDLC Integration: Study real-world exploitability and integration into the Software Development Life Cycle.
• Dynamic Vulnerability Detection: Explore AI-based methods for dynamic vulnerability detection.