Data Poisoning and Shilling in Recommender Systems

535 concepts232 papers1 proposal

Recommender systems operate on a fundamental assumption: user interactions reflect genuine preferences. Collaborative filtering algorithms map these interactions into a shared latent space where proximity equals relevance. If a user likes items $A$ and $B$, and another user likes item $A$, the system infers that the second user might also enjoy item $B$.

Shilling attacks (or data poisoning) exploit this foundational assumption. By injecting a small set of carefully crafted fake user profiles into the training data, an adversary can warp this latent space, artificially drawing a target item closer to a specific demographic or the entire user base. Understanding why these attacks work requires looking past the surface-level mechanics of "fake accounts" and examining how recommendation algorithms mathematically process similarity, influence, and optimization.

From Heuristics to Gradient-Driven Profiles

Early shilling strategies relied on simple heuristics. An attacker would create profiles containing the target item alongside a selection of "filler" items—either chosen at random or selected because they had high average ratings across the platform, as cataloged by Hofmann et al. in A Survey of Attack-Resistant Collaborative Filtering Algorithms.

The intuition was straightforward: by co-rating the target item with popular items, the fake profiles would appear as "neighbors" to a vast number of real users in memory-based algorithms like $k$-Nearest Neighbors ($k$-NN). However, these heuristic attacks are noisy and easily neutralized. As shown by Mobasher et al. in Model-Based Collaborative Filtering as a Defense against Profile Injection Attacks, model-based approaches like Probabilistic Latent Semantic Analysis (PLSA) smooth out this random noise by anchoring recommendations in global latent structures rather than raw user-to-user matches.

To bypass these defenses, modern attacks treat data poisoning as a bi-level optimization problem. As demonstrated by Li et al. in Data Poisoning Attacks on Factorization-Based Collaborative Filtering, attackers can leverage Karush-Kuhn-Tucker (KKT) conditions to calculate exactly how a matrix factorization model's parameters will shift in response to new data. Instead of guessing filler items, gradient-based optimization crafts synthetic profiles that perfectly mimic the statistical distributions of legitimate users while maximizing the predicted rating of the target item.

The Concentration of Influence

Why can a tiny fraction of fake users (often between 1% and 3% of the user base) cause massive, system-wide shifts in recommendations? The vulnerability stems from a high concentration of influence within latent factor models.

Cheng et al. explored this phenomenon through Fast Influence Analysis in Incorporating Interpretability into Latent Factor Models via Fast Influence Analysis, demonstrating that a model's final recommendations are often driven by a remarkably small fraction of training points. If an attacker can identify or generate these highly influential interactions, they can exert outsized control over the system's outputs.

[Legitimate Users] ----\
                        +---> [Latent Factor Space] ---> [Skewed Top-K Recommendations]
[1% Optimized Bots] ----/            ^
                                     | (Warped by high-influence points)

This concentration of influence is particularly severe for long-tail, unpopular items. As Fang et al. demonstrated in Poisoning Attacks to Graph-Based Recommender Systems, injecting just 1% of fake users into a graph-based recommender can amplify an unpopular item's recommendation rate by up to 580 times.

Because unpopular items have a near-zero baseline of existing interactions, their positions within the latent space or interaction graph are highly volatile. A popular item is anchored in place by thousands of genuine connections, but an unpopular item's representation can be completely dragged across the latent space by only a handful of coordinated adversarial profiles, as noted by Anelli et al. in Adversarial Attacks against Visual Recommendation: an Investigation on the Influence of Items' Popularity.

The Paradox of Advanced Architectures

One might expect advanced architectures—such as deep neural networks or contrastive learning frameworks—to be inherently more robust against data poisoning. However, the structural mechanisms designed to improve recommendation quality often introduce unique security vulnerabilities.

Consider contrastive learning, which is frequently used to mitigate popularity bias by forcing the model to learn evenly dispersed item embeddings. Wang et al. revealed a striking paradox in Unveiling Vulnerabilities of Contrastive Recommender Systems to Poisoning Attacks: the very repulsion forces used to break down dense clusters of popular items actually lower the optimization barrier for attackers. By dispersing the embeddings, the system reduces the mathematical effort required for an adversarial framework to push a target item into a user's top-K recommendation list.

Furthermore, as deep learning models allow for more complex user behavior modeling, neural network-based attacks can generate synthetic profiles that bypass traditional statistical anomaly detectors by blending seamlessly into genuine user distributions, a challenge highlighted by Ye et al. in RecAD: Towards A Unified Library for Recommender Attack and Defense.

Expanding to New Modalities: Text and Sequences

As recommendation systems have evolved past simple user-item rating matrices, poisoning attacks have adapted to exploit semantic and temporal dependencies.

Semantic Manipulation

In Review-Based Recommender Systems (RBRS) and ID-free systems, recommendations are driven by textual features rather than interaction IDs. Chiang et al. showed in Shilling Black-box Review-based Recommender Systems through Fake Review Generation that language models can be optimized to generate fake reviews containing specific semantic structures and keywords. These linguistic patterns trick the text-profiling components into aligning a low-quality product with the profile of high-performing items, bypassing the need to manipulate numerical rating matrices entirely.

Profile Pollution

In sequential recommender systems, which predict a user's next action based on the chronological order of their history, attackers do not even need to create fake accounts. Instead, they can execute a highly localized attack known as profile pollution.

As analyzed by Yue et al. in Defending Substitution-Based Profile Pollution Attacks on Sequential Recommenders, substituting just one or two items in a legitimate user's active session can catastrophically disrupt their recommendation sequence. By exploiting the tight temporal dependencies learned by recurrent neural networks or transformers, the attacker forces the model to pivot its immediate next-item predictions toward the target product.

Key Foundational Work

Recommender Systems: Attack Types and Strategies (2005) — contributed 12 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 241 papers
  2. Version 1 · 2026-10-04 · 281 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_shilling_attacks_a5d9f344,
  title = {shilling attacks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/shilling-attacks/rdr_a5d9f3445fee4a289f470dd335ad320a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_shilling_attacks_a5d9f344_v2,
  title = {shilling attacks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/shilling-attacks/rdr_a5d9f3445fee4a289f470dd335ad320a/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake