Secure and scalable client pre-filtering and authentication framework for federated learning systems

Abstract

Abstract A federated learning system, which emphasises decentralisation, allows cooperative model training to ensure the privacy and security of the data. Client authentication is an important factor in a decentralised system to prevent unauthorised access and unauthorised users from adding exceptions or irrelevant facts during model preparation. Due to which model becomes weak, poisoned, and inconsistent. The proposed framework adopts a two-stage client verification pipeline. In the first stage, probabilistic data structures serve as a lightweight admission pre-filter that screens candidate clients before any cryptographic operation is performed: a Bloom Filter (BF) is used for static client populations, and an Adaptive Quotient Filter (AQF) is used for dynamic client populations, enabling fast membership screening both before base-model distribution and during collection of gradient updates. In the second stage, clients admitted by the pre-filter undergo ECDSA-based signature verification, which constitutes the actual authentication step and compensates for the residual false positives inherent to probabilistic filtering. This design allows the server to reject the majority of unauthorized requests cheaply via BF/AQF while reserving expensive cryptographic verification for genuine candidates. The dataset used in this study is the IP Address Geographic Profiles dataset, comprising 129,857 entries. Evaluation covered memory utilisation, lookup time, and false-positive behaviour of BF and AQF against six alternative data structures. BF and AQF reduced server-side memory usage by up to 90% relative to a Trie (e.g., 1.9 MB and 3 MB versus 23.2 MB for 100,000 password-derived credentials), at the cost of a modest lookup-time overhead relative to hash-based structures (0.054–0.064 s versus 0.003–0.006 s for a HashSet on 100,000 records). At the protocol level, combining BF/AQF pre-filtering with ECDSA verification reduced authentication messages and communication cost by roughly 30% relative to ECDSA-only authentication under a simulated environment with 30% Byzantine clients, with the gain compounding over rounds (communication cost falling from 125,000 KB to 87,537.5 KB over 50 rounds). The study concluded that the proposed model provides a secure and scalable client screening and authentication framework by combining BF/AQF with ECDSA-based signature verification in a decentralised environment where multiple clients participate from different regions. Experimental results present significant changes in memory usage, performance analysis, and resistance to unauthorised clients.