Trustless Federated Learning at Edge-Scale: A Compositional Architecture for Decentralized, Verifiable, and Incentive-Aligned Coordination

TL;DR

This work tackles the trust gap in edge federated learning by proposing a compositional architecture that enables verifiable, incentive-aligned training at edge scale. It introduces four mechanisms—SumIntegrityProofs for verifiable weighted aggregation, geometric novelty decomposition to deter replay and Sybil attacks, object-centric coordination for O(N) parallel updates, and time-locked governance to prevent retroactive policy manipulation. The security analysis reduces to established primitives (Secure Aggregation, differential privacy, and Byzantine fault-tolerant consensus) and is complemented by cost and scalability projections for 10,000 contributors and 20M parameters. If validated on a testnet, this approach could deliver cryptographic accountability and economically sustainable, edge-native learning for billions of devices.

Abstract

Artificial intelligence is retracing the Internet's path from centralized provision to distributed creation. Initially, resource-intensive computation concentrates within institutions capable of training and serving large models.Eventually, as federated learning matures, billions of edge devices holding sensitive data will be able to collectively improve models without surrendering raw information, enabling both contribution and consumption at scale. This democratic vision remains unrealized due to certain compositional gaps; aggregators handle updates without accountability, economic mechanisms are lacking and even when present remain vulnerable to gaming, coordination serializes state modifications limiting scalability, and governance permits retroactive manipulation. This work addresses these gaps by leveraging cryptographic receipts to prove aggregation correctness, geometric novelty measurement to prevent incentive gaming, parallel object ownership to achieve linear scalability, and time-locked policies to check retroactive manipulation.

Figures (8)

• Figure 1: Evolution of AI architectural paradigms: from centralized cloud training, to distributed inference with centralized training, to blockchain-coordinated edge-native collaborative learning.
• Figure 2: Geometric novelty decomposition for replay-resistant contribution measurement. Novel directions (perpendicular to basis memory) receive rewards, while repeated or redundant updates (parallel to explored directions) yield zero novelty score.
• Figure 3: Object-centric coordination architecture decomposing state into parallel owned ContributorRegistry objects and infrequent shared coordination layer, enabling $O(N)$ scaling versus traditional $O(N^2)$ serialization.
• Figure 4: Separation of concerns across four operational planes: Control (round lifecycle), Data (secure aggregation), Incentive (reward distribution), and Audit (cryptographic verification). Owned objects (dark) enable parallel updates; shared objects (light) require consensus.
• Figure 5: Complete system architecture showing six participant classes and their interactions. Contributors upload masked adapters to decentralized storage and submit CIDs to blockchain; committee validators fetch adapters, perform aggregation, and write results back to storage; receivers escrow fees and fetch trained models; auditors verify cryptographic receipts independently.