Incentives and Contribution Evaluation in Secure Collaborative Learning

Collaborative machine learning, most prominently Federated Learning (FL), is increasingly deployed in security- and privacy-critical settings, from healthcare to fraud detection, without sharing raw data. The security community has extensively studied how to protect such systems: secure aggregation, differential privacy, and robust aggregation defend against curious servers, inference attacks, and poisoning. Far less attention has been paid to a component that is simultaneously a security mechanism and an attack surface: contribution evaluation — quantifying how much each participant’s data and updates improve the jointly trained model. As a security mechanism, contribution evaluation underpins free-rider and misbehavior detection, accountability, and auditability, providing the technical basis for spotting low-quality, stale, or poisoned updates. As an attack surface, contribution scores are themselves a target: they can be manipulated by strategic participants who inflate their own scores or suppress honest contributors, they are fragile to the choice of aggregation rule, and their computation can leak sensitive information. Principled game-theoretic measures such as the Shapley value are computationally prohibitive and often incompatible with the very privacy-enhancing technologies that make collaborative learning deployable. The evaluation pipeline, from score approximation to reward settlement, must therefore be secure, verifiable, and auditable — especially in cross-device and edge federations with resource-constrained, intermittently available participants. These challenges sharpen as collaboration extends beyond training a shared model towards multi-agent systems, in which autonomous, increasingly LLM-driven agents owned by different parties interact to solve a common task: credit must be assigned over long, ordered interaction traces, agents may misreport their competence or free-ride on others’ reasoning, and contribution signals must be computed without exposing private data or prompts. This Research Topic aims to establish contribution evaluation and incentives as a first-class concern of secure machine learning, consolidating a landscape scattered across security, distributed ML and edge systems, and game theory venues. Substantially extended versions of papers from a corresponding workshop are welcome alongside original submissions, and the collection is open to all contributions within scope. We welcome original research, systematization-of-knowledge, survey, and perspective articles on topics including: - attacks on contribution evaluation — gaming, collusion, Sybil behavior, score inflation, poisoning — and defenses; - manipulation-resistant and Byzantine-robust scoring; - contribution evaluation under privacy-enhancing technologies; - privacy leakage of contribution scores; - free-riding, misbehavior detection, and accountability; - secure, verifiable reward distribution, including blockchain- and smart-contract-based mechanisms; - cost-aware incentives and client selection under heterogeneity, energy, and availability constraints; - mechanism design for secure collaborative learning; - fairness of reward allocation; - evaluation metrics, axiomatic foundations, and scalable game-theoretic approximations; - data valuation, marketplaces, and the pricing of data and privacy; - participation, coalition formation, and interdependent privacy; - credit assignment and incentives in multi-agent and LLM-agent collaboration; and - benchmarks and deployment experiences from real-world cross-silo and cross-device federations. By assembling these threads in a single open-access collection, the Research Topic aims to close a critical gap in the security of collaborative machine learning and to shape a shared research agenda at the intersection of computer security, incentives, and distributed learning. Topic editors: Balázs Pejó, Steve Drew, Qiang Tang, Delio Jaramillo-Velez Source: https://www.frontiersin.org/research-topics/85885/incentives-and-contribution-evaluation-in-secure-collaborative-learning