TL;DR
This paper introduces IndirectAD, a novel data poisoning attack against recommender systems that significantly lowers the required attack cost by using a "trigger item" to indirectly promote a target item. This method achieves notable impact with as little as 0.05% of user data, highlighting a more realistic and potent threat to recommender system security.
What Problem Does This Paper Address?
Recommender systems are ubiquitous, personalizing content across e-commerce, streaming, and social media. While beneficial, they are susceptible to data poisoning attacks, where malicious actors inject fake data to manipulate recommendations. Traditionally, these attacks required controlling a substantial portion of user data (e.g., >1%) to be effective, making them impractical for large platforms. This paper addresses the feasibility gap, demonstrating how to achieve significant recommendation manipulation with a drastically reduced attack cost, thereby posing a more realistic threat.
Methods
The core of IndirectAD is inspired by Trojan attacks in machine learning. Instead of directly promoting a target item, which might be unpopular or irrelevant to users, IndirectAD employs a two-pronged strategy:
- Trigger Item Selection: The attacker first identifies an item with broad appeal ("trigger item") that is easier to promote. This is done by evaluating candidate items based on how much their adversarial loss decreases after a single optimization step.
- Indirect Promotion: The attacker then creates fake user profiles that interact with both the trigger item and the desired target item. This establishes a co-occurrence signal. Through adversarial optimization, the attacker fine-tunes these interactions to strengthen the association between the trigger and target items. This leverages the recommender system's tendency to recommend items that are frequently consumed together, indirectly transferring the promotion effect from the trigger to the target item.
The attack uses a substitute model trained on partial data to approximate the target recommender system's behavior and employs Projected Gradient Descent (PGD) for adversarial optimization.
Datasets
The authors evaluated their approach on four real-world datasets:
- Amazon Books [15]: User interactions with books on Amazon.
- Amazon Arts [15]: User interactions with arts and crafts products on Amazon.
- Steam [3]: User activity and interactions on the Steam gaming platform.
- ML-1M [16]: Movie ratings from 6,000 users on 4,000 movies.
Key Insights and Ideas
- The Power of Indirectness: The most significant insight is that promoting a "trigger item" that users are more likely to engage with can be a more effective strategy than directly pushing a less desirable target item. This indirect approach bypasses the need for direct user interest alignment.
- Reduced Attack Cost: The core contribution is demonstrating that effective data poisoning can be achieved with extremely low poisoning ratios (as low as 0.05%), making it a practical threat even for large-scale platforms.
- Leveraging Co-occurrence: Recommender systems' reliance on item co-occurrence for making predictions is exploited as a vulnerability. By artificially creating co-occurrences, attackers can manipulate these learned associations.
- Transferability: The attack shows good transferability, meaning a substitute model trained on one type of recommender system can be used to attack others, even with different architectures.
Key Results
- Significant Impact at Low Ratios: IndirectAD achieved a noticeable impact on recommendation quality (measured by Hit Ratio at 20, HR@20) with only 0.05% of the user base. For example, on Amazon Books with WRMF, it achieved a 0.4762% HR@20 increase, while a state-of-the-art attack failed to show any improvement.
- Outperforming Baselines: IndirectAD consistently outperformed the state-of-the-art injection attack method [13] across various datasets and recommender systems, often by a significant margin (e.g., eightfold increase in HR@20).
- Effectiveness Across Item Popularity: The attack proved effective even for items with low popularity ("tail" items), where traditional attacks struggled or failed.
- Robustness to Defenses: The attack demonstrated low detectability by existing defense mechanisms, with an AUC score of only 0.22 for the "Catch the Black Sheep" method.
Limitations and Future Work
- Dataset Size for Low Ratios: The paper notes that for very small datasets like ML-1M, a 0.01% poisoning ratio might involve fewer than one user, making it experimentally infeasible.
- User Diversity Impact: While the attack is generally effective, its performance is slightly less pronounced in random user group settings compared to clustered groups, suggesting that more unified user characteristics can be more easily exploited.
- Temporal Dynamics: The authors suggest that incorporating temporal dynamics into the attack strategy could further enhance its effectiveness, as user behavior often follows temporal patterns. They propose reinforcement learning as a potential method for optimizing temporal attack strategies.
- Direct Marketing: The paper mentions that direct marketing channels (e.g., promotional emails) could be used to boost trigger item visibility, complementing the recommender system attack.
Summary updated