Introduction
Sequential Recommender Systems (SRSs) have become the backbone of modern digital platforms, predicting user preferences based on chronological interaction histories. However, their reliance on sequential patterns introduces a specific vulnerability: profile pollution. Unlike traditional data poisoning, which involves injecting fake accounts into training sets, profile pollution involves manipulating a legitimate user's history—often via unauthorized access or cross-site request forgery—to degrade or redirect recommendations. This paper investigates a particularly stealthy form of this attack—item substitution—and proposes robust defense mechanisms to mitigate its impact.
The Challenge of Discrete Profile Pollution
In domains like computer vision, adversarial attacks often involve adding imperceptible continuous noise to pixels. Recommender systems, however, operate in a discrete space. One cannot add "fractional" items to a user's history; instead, an attacker must swap existing items for others.
The authors introduce a Gradient-Guided Substitution attack to demonstrate this vulnerability. By analyzing the model's gradients, the attack identifies "vulnerable" items—those with the highest gradient norms—and replaces them with adversarial alternatives. To maintain stealth and bypass anomaly detection, the replacement items are constrained to be semantically similar to the originals within the embedding space. This ensures the polluted profile remains plausible while significantly damaging recommendation accuracy.
Proposed Defense Mechanisms
To counter these targeted manipulations, the researchers propose two primary defense strategies designed to "smooth" the model's decision boundaries and reduce sensitivity to discrete swaps.
1. Dirichlet Neighborhood Sampling (DNS)
DNS is a randomized training technique that moves away from point-wise item representations. Instead of training on a fixed item $i$, the model considers a "neighborhood" of semantically similar items. During training, a new embedding is sampled from the convex hull of this neighborhood using a Dirichlet distribution. This process forces the model to learn that minor substitutions within a neighborhood should not lead to drastic changes in the final recommendation, effectively building resilience against item swaps.
2. Adversarial Training with Mixed Representations (AdvTrain)
AdvTrain employs a minimax optimization strategy. During the training phase, the model actively searches for the "worst-case" linear combination of item embeddings that would maximize the recommendation loss. The model parameters are then updated to minimize this worst-case loss. By training on these "mixed representations," the model becomes robust to a wider range of potential perturbations compared to simple neighborhood sampling.
Experimental Evaluation
The effectiveness of these methods was evaluated across five diverse datasets: Amazon Beauty, MovieLens (1M & 20M), LastFM, and Steam. The study tested four representative architectures: NARM (RNN-based), SASRec (Transformer-based), BERT4Rec, and Locker.
Key Insights
- Attack Potency: The substitution attack proved highly effective, causing a drop of over 50% in NDCG@10 on the MovieLens 1M dataset in untargeted scenarios.
- Defense Efficacy: AdvTrain emerged as the most robust defense, outperforming other methods in 30 out of 40 tested scenarios. The proposed defenses reduced performance variations under attack by over 50% across multiple datasets.
- Architecture Matters: Transformer-based models (SASRec, BERT4Rec) exhibited higher natural robustness than RNN-based models (NARM). This suggests that global attention mechanisms, which distribute influence across the entire sequence, are less susceptible to single-point failures than the hidden state bottlenecks in RNNs.
- Minimal Overhead: Crucially, these defenses provided significant security gains with very limited impact on "clean" performance, meaning accuracy is not sacrificed for security.
Conclusion and Future Work
This research highlights the critical need for security in sequential recommendation. While the proposed defenses are effective, several avenues for future work remain. The current attack model assumes a "white-box" scenario where the attacker has access to model weights; exploring black-box attacks is a logical next step. Additionally, the authors suggest moving toward certified robustness using techniques like Interval Bound Propagation to provide mathematical guarantees against profile pollution. Extending these defenses to graph-based or general collaborative filtering models remains a significant open challenge.
Summary updated