Research map / Reinforcement learning / Visual and representation RL

BertRLFuzzer: A BERT and Reinforcement Learning Based Fuzzer (Student Abstract)

AAAI 2024 · Reinforcement learning › Visual and representation RL · approach: RL with tree search for discovery · 7 citations (OpenAlex)

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.

Presents BertRLFuzzer, a fuzzer that uses a BERT model as the actor in a PPO reinforcement-learning loop, with a multi-armed bandit seed filter, to generate grammar-adhering attack vectors against web applications.

Problem

Finding security vulnerabilities (e.g., SQL injection and XSS) in web applications with grammar-adhering fuzzing, which is time-consuming to build by hand and must be redone when new vulnerability classes appear.

Approach

Pre-train a BERT model on attack-vector seed inputs so it learns the input grammar, then fine-tune it as the actor-critic PPO agent in an RL loop driven by victim-application rewards and penalties, with a multi-armed bandit choosing seed inputs.

Key results

• BertRLFuzzer reaches first attack in 102 seconds, 54% faster than the nearest competing tool (DeepXSS at 224 s in Table 1 is listed as nearest competitor per the text).
• Finds 61 vulnerabilities versus 44 for DeepFix and 40 for DeepXSS, i.e. 17 new vulnerabilities over the nearest competitor.
• Using BERT as the RL agent reduces parser penalties by 13% relative to a vanilla RL agent (Table 2 ablation).

Measured effects reported in the paper

Surprising findings

• The hand-written grammar-adhering mutation fuzzer has 0% parser penalty, beating BertRLFuzzer's 5%, which the authors attribute to explicitly specified grammar.

Limitations

• Parser penalty is higher than the handwritten grammar-adhering mutator (5% vs 0%).
• Grammar adherence is only probabilistic ('likely grammar-adhering'), not guaranteed.
• Requires a set of known-class seed inputs representative of the target attack grammar.

Experimental setup

Datasets: over 7K injection attack vectors from public GitHub repositories

Benchmarks: benchmark of 9 victim websites (over 16K LOC)

Metrics: time to first attack unique fields vulnerabilities found parser penalties

Baselines: DeepFix DeepXSS DeepFuzz DQN fuzzer Baseline Grammar Mutator Baseline Random Mutator Baseline Random Fuzzer

Builds on: BERT PPO DeepFuzz DeepXSS DeepFix

Closest papers by meaning

• Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use (ICML 2026 · similarity 0.84)
• Learning Diverse Attacks on Large Language Models for Robust Red-Teaming and Safety Tuning (ICLR 2025 · similarity 0.83)
• BadRL: Sparse Targeted Backdoor Attack against Reinforcement Learning (AAAI 2024 · similarity 0.83)
• Nemotron-Research-Tool-N1: Exploring Tool-Using Language Models with Reinforced Reasoning (ICLR 2026 · similarity 0.81)
• Text-Diffusion Red-Teaming of Large Language Models: Unveiling Harmful Behaviors with Proximity Constraints (AAAI 2025 · similarity 0.81)
• Certifying Language Model Robustness with Fuzzed Randomized Smoothing: An Efficient Defense Against Backdoor Attacks (ICLR 2025 · similarity 0.81)