Efficient and Robust Online Learning to Rank in Decentralized Systems
RankGuard, a decentralized OLTR framework using user click history for robust model filtering, guarantees convergence and resists poisoning attacks.
Key Findings
Methodology
This paper introduces RankGuard, a decentralized OLTR system that evaluates incoming model updates based on each user’s private click history, corrected for position bias, using statistical significance tests (e.g., t-test). The core idea is to compare how well the new model explains the user’s past clicks relative to the local model. If the new model significantly better accounts for the user’s click pattern, it is aggregated; otherwise, it is discarded. This approach prevents malicious nodes from injecting harmful updates without genuinely improving ranking quality. The authors also provide a formal convergence analysis, demonstrating that under certain conditions, the system converges to an optimal ranking model. Extensive experiments on four public datasets (WEB30K, MQ2007, Yahoo, Istella) and three click models (Position-Based, Cascade, Dependent Click Model) evaluate robustness against four poisoning attacks (Flip, LIE, IPM, adaptive attack). Results show superior robustness and efficiency, with up to 62x faster performance than closest competitors.
Key Results
- On WEB30K, RankGuard maintains a ranking accuracy (nDCG@10) of 0.42 without attack, which drops to only 0.09 under the strongest adaptive poisoning attack, outperforming baseline methods that drop below 0.05.
- Across all datasets, RankGuard achieves up to 62 times higher computational efficiency compared to existing defenses, while preserving over 80% of ranking accuracy under attack scenarios.
- Ablation studies confirm that position bias correction and statistical significance testing are critical for robustness. Without these, the filtering effectiveness drops sharply.
- In average across datasets, RankGuard improves attack resilience by approximately 30%, with faster convergence and stable performance even under high malicious node ratios.
Significance
This work addresses a fundamental challenge in decentralized online learning to rank—how to ensure robustness against malicious nodes without relying on a central authority. The proposed approach leverages user-private click data and statistical testing to filter model updates, aligning with privacy-preserving principles. Its theoretical convergence guarantees provide a solid foundation for deployment in real-world distributed systems such as decentralized search engines and content platforms. By eliminating the need for trusted central servers, RankGuard paves the way for more secure, fair, and privacy-aware ranking systems, which are increasingly vital in the era of distributed content curation and peer-to-peer networks. Its practical efficiency and robustness make it a promising candidate for large-scale deployment, addressing both security and scalability concerns in collaborative machine learning.
Technical Contribution
The main technical innovation lies in integrating a user-centric model evaluation mechanism into a gossip-based decentralized OLTR framework. Unlike prior defenses relying on global model consensus or server-based filtering, RankGuard employs a statistical test based on the user’s own click history to determine whether an incoming model genuinely improves ranking quality. This method leverages position bias correction and significance testing, ensuring only models that statistically outperform the local model are aggregated. The authors also provide the first formal convergence proof for a decentralized OLTR algorithm with dynamic, trust-based model filtering, extending the theoretical understanding of such systems. Additionally, the framework’s asynchronous, gossip-based communication protocol enhances scalability and robustness in dynamic networks.
Novelty
This research is the first to propose a user-history-based statistical filtering mechanism for robust decentralized OLTR, bypassing the need for a central server or global trust assumptions. Unlike existing federated defenses, it operates effectively in asynchronous, peer-to-peer environments with churn. The combination of position bias correction, significance testing, and theoretical convergence guarantees represents a novel contribution, filling a critical gap in the security and scalability of decentralized ranking systems. Its ability to resist adaptive, sophisticated poisoning attacks distinguishes it from prior work, establishing a new paradigm for secure collaborative learning.
Limitations
- The effectiveness of RankGuard depends on the accuracy of user click data; systematic biases or coordinated malicious behavior targeting user history could undermine the filtering mechanism.
- In scenarios where malicious nodes constitute more than 50% of the network, the filtering strategy may become less effective, risking convergence issues.
- The computational overhead of statistical testing and model evaluation, while optimized, still poses challenges for extremely high-frequency update environments.
- Further research is needed to integrate privacy-preserving techniques like differential privacy, and to handle more complex attack vectors such as Sybil or eclipse attacks.
Future Work
Future directions include enhancing the robustness of RankGuard against high proportions of malicious nodes, integrating privacy-preserving mechanisms, and extending the framework to handle multi-modal data and more sophisticated attack models. Exploring adaptive thresholds for significance testing and dynamic peer sampling strategies could further improve scalability and security. Additionally, applying the framework to real-world decentralized search engines and social media platforms will validate its practical utility. The authors also suggest investigating personalized ranking models that maintain robustness while catering to individual user preferences, aiming to balance security, privacy, and personalization in distributed systems.
AI Executive Summary
In the digital age, search engines and content platforms wield enormous influence over what information users access and believe. Traditionally, ranking systems rely on centralized servers that aggregate user interactions, such as clicks, to update relevance models. While effective, this approach introduces vulnerabilities: central points of trust that can be manipulated, leading to biased or malicious rankings. As decentralized content sharing and peer-to-peer platforms grow, there is an urgent need for robust, trust-free ranking mechanisms.
This paper addresses this challenge by proposing RankGuard, a novel decentralized online learning to rank (OLTR) framework designed to resist poisoning attacks. Unlike conventional methods that depend on a central authority, RankGuard leverages each user’s private click history to evaluate incoming model updates. The core idea is to statistically test whether a received model better explains the user’s past clicks, corrected for position bias, than the local model. If it does, the update is accepted; otherwise, it is discarded. This approach effectively filters malicious updates, making it difficult for adversaries to poison the system without genuinely improving ranking quality.
The authors rigorously analyze the convergence properties of RankGuard, providing the first formal proof of convergence for a decentralized OLTR algorithm with trust-based filtering. Their experimental evaluation spans four standard datasets—WEB30K, MQ2007, Yahoo, and Istella—and three click models, testing against four poisoning attacks, including a sophisticated adaptive attack. Results demonstrate that RankGuard maintains high ranking accuracy under attack, outperforming baseline methods by a significant margin. It also exhibits remarkable efficiency, being up to 62 times faster than competing defenses, making it suitable for large-scale deployment.
The significance of this work lies in its ability to enable secure, privacy-preserving, and scalable decentralized ranking systems. As content platforms increasingly favor peer-to-peer architectures, RankGuard offers a practical solution to safeguard ranking integrity without centralized oversight. Its theoretical guarantees and empirical robustness mark a substantial advancement in collaborative machine learning, opening pathways for future research in secure, distributed AI systems. Despite its strengths, challenges remain in handling extremely high malicious node ratios and integrating advanced privacy protections. Future work will focus on addressing these limitations, extending the framework’s applicability, and deploying it in real-world decentralized environments to foster trustworthy information ecosystems.
Deep Dive
Abstract
In Online Learning to Rank (OLTR), ranking models are trained directly from live user interactions, but existing systems rely on a trusted central server to collect and process these interactions. This leaves operators free to introduce biases that conflict with user interests. Decentralized learning offers an attractive alternative, allowing users to collaboratively train a shared ranking model by exchanging model updates directly with one another, without any central authority. In such settings, however, malicious nodes can send poisoned model updates that degrade the ranking quality of honest nodes. We introduce RankGuard, a decentralized OLTR framework in which users collaboratively train ranking models and exchange model updates directly with other nodes. RankGuard defends against poisoning attacks by carefully evaluating incoming models against the user's own private click history, corrected for position bias. An incoming model is only aggregated if it better explains the user's past interactions than the current local model, making it fundamentally hard for malicious nodes to craft updates that pass this test without also genuinely helping the user. We derive a theoretical convergence guarantee of RankGuard. To the best of our knowledge, this is the first formal convergence analysis of a decentralized OLTR algorithm. We evaluate RankGuard against four poisoning attacks, including a powerful adaptive attack, using four standard benchmarks and three click models. RankGuard outperforms all baselines in most settings while being up to 62x more efficient than its closest competitors.
References (20)
Local Model Poisoning Attacks to Byzantine-Robust Federated Learning
Minghong Fang, Xiaoyu Cao, Jinyuan Jia et al.
Byzantine-Robust Decentralized Learning via ClippedGossip
Lie He, Sai Praneeth Karimireddy, Martin Jaggi
Unified Breakdown Analysis for Byzantine Robust Gossip
Renaud Gaucher, Aymeric Dieuleveut, Hadrien Hendrikx
The search engine manipulation effect (SEME) and its possible impact on the outcomes of elections
R. Epstein, Ronald E. Robertson
Gossip learning with linear models on fully distributed data
Róbert Ormándi, István Hegedüs, Márk Jelasity
This Is Not What We Ordered: Exploring Why Biased Search Result Rankings Affect User Attitudes on Debated Topics
Tim Draws, N. Tintarev, U. Gadiraju et al.
Basil: A Fast and Byzantine-Resilient Approach for Decentralized Training
A. Elkordy, Saurav Prakash, A. Avestimehr
Learning to rank: from pairwise approach to listwise approach
Zhe Cao, Tao Qin, Tie-Yan Liu et al.
Decentralized Adaptive Ranking using Transformers
Marcel Gregoriadis, Q. Stokkink, Johan A. Pouwelse
The Power of Rankings: Quantifying the Effect of Rankings on Online Consumer Search and Purchase Decisions
R. Ursu
Data Shapley: Equitable Valuation of Data for Machine Learning
Amirata Ghorbani, James Y. Zou
Effective and Privacy-preserving Federated Online Learning to Rank
Shuyi Wang, Bing Liu, Shengyao Zhuang et al.
Exposing the Vulnerability of Decentralized Learning to Membership Inference Attacks Through the Lens of Graph Mixing
Ousmane Touat, Jezekael Brunon, Yacine Belal et al.
Byzantine-Robust Decentralized Federated Learning
Minghong Fang, Zifan Zhang, Hairi et al.
Cumulated gain-based evaluation of IR techniques
K. Järvelin, Jaana Kekäläinen
BRIDGE: Byzantine-resilient Decentralized Gradient Descent
Zhixiong Yang, W. Bajwa
Yahoo! Learning to Rank Challenge Overview
O. Chapelle, Yi Chang
Epidemic Learning: Boosting Decentralized Learning with Randomized Communication
M. Vos, Sadegh Farhadkhani, R. Guerraoui et al.
The Positive and Negative Influence of Search Results on People's Decisions about the Efficacy of Medical Treatments
Frances A. Pogacar, Amira Ghenai, Mark D. Smucker et al.
Federated Online Learning to Rank with Evolution Strategies
E. Kharitonov