Online Learning to Rank: Robustness and Adversarial Attacks

Overview

This project investigates the robustness of recommender systems through online learning and multi-armed bandit (MAB) algorithms, with a particular focus on adversarial attacks and defense.

Working under Professor Jinhang Zuo along with Eric He and Qirun Zheng, this research explores how knowledge from adversarial attacks on stochastic bandits can be leveraged to develop both:

• Novel attack algorithms targeting recommender systems
• More robust reinforcement learning (RL) policies to defend against such attacks

We develop a method to attack bandits via fake data injection.

---

Project Structure

online_learning_torank/
├── algorithms/ # Bandit and ranking algorithms
├── dataset/ # Dataset files and preprocessing
├── thompson/ # Thompson sampling implementations
├── observation_free/ # Observation-free attack code
├── attack.py # Core attack logic
├── attack_real.py # Real-world attack experiments
├── attack_log.py # Attack logging utilities
├── movielens.py # MovieLens dataset integration
├── params_experiment.py # Experiment configuration
├── figures/ # Plots and visualizations
└── dataset.txt # Dataset details

Running Experiments

python attack_real.py

Research

arXiv 05.2025: Practical Adversarial Attacks on Stochastic Bandits via Fake Data Injection