{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,2,21]],"date-time":"2026-02-21T14:03:34Z","timestamp":1771682614375,"version":"3.50.1"},"reference-count":0,"publisher":"Slovenian Association Informatika","issue":"7","content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":["IJCAI"],"abstract":"<jats:p>Federated Learning (FL) enables decentralized model training by having clients process local data andtransmit only learned updates to a central server, but faces significant communication bottlenecks due tofrequent exchanges of high-dimensional model parameters. While existing compression techniques suchas sparsification and quantization reduce overhead, they fail to fully exploit the structural properties offloating-point representations. This paper introduces a novel compression strategy that innovativelyexploits the exponent field of float16 representations to encode sequences of negligible parameter updates,achieving substantial communication reduction while preserving model integrity. The proposed methodoperates through four main steps: (1) client-side model subtraction between local and global parameters,(2) downcasting to float16, (3) threshold-based pruning to remove insignificant values, and (4) exponentbased encoding to compactly represent sequences of negligible updates, followed by server-sidedecompression through exponent extraction and sequence reconstruction. On MNIST, the methodachieves an 88.4% size reduction (threshold 0.001) with merely 0.2% accuracy loss versus baseline, whilea lighter threshold (0.0001) improves accuracy by 0.1% at 64.7% compression; on CIFAR-10, it yields64% compression with maintained accuracy (+0.1%) and 42% faster convergence through critical weightpreservation, with lighter thresholds (0.0001) achieving 51.4% size reduction and 1.2% accuracyimprovement. The technique\u2019s computational efficiency and compatibility with existing FL frameworksmake it particularly suitable for resource-constrained edge environments, bridging a critical gap incommunication-efficient FL through innovative use of floating-point exponent manipulation for scalable,privacy-preserving distributed learning.<\/jats:p>","DOI":"10.31449\/inf.v50i7.8833","type":"journal-article","created":{"date-parts":[[2026,2,21]],"date-time":"2026-02-21T13:16:16Z","timestamp":1771679776000},"source":"Crossref","is-referenced-by-count":0,"title":["Exploiting the Exponent of Floating-Point: A Novel Pathway to Efficient Federated Learning"],"prefix":"10.31449","volume":"50","author":[{"given":"Goran Saman","family":"Nariman","sequence":"first","affiliation":[],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Hozan Khalid","family":"Hamarashid","sequence":"additional","affiliation":[],"role":[{"role":"author","vocabulary":"crossref"}]}],"member":"16141","published-online":{"date-parts":[[2026,2,21]]},"container-title":["Informatica"],"original-title":[],"link":[{"URL":"https:\/\/www.informatica.si\/index.php\/informatica\/article\/download\/8833\/6491","content-type":"application\/pdf","content-version":"vor","intended-application":"text-mining"},{"URL":"https:\/\/www.informatica.si\/index.php\/informatica\/article\/download\/8833\/6491","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,2,21]],"date-time":"2026-02-21T13:16:16Z","timestamp":1771679776000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.informatica.si\/index.php\/informatica\/article\/view\/8833"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,2,21]]},"references-count":0,"journal-issue":{"issue":"7","published-online":{"date-parts":[[2026,2,21]]}},"URL":"https:\/\/doi.org\/10.31449\/inf.v50i7.8833","relation":{},"ISSN":["1854-3871","0350-5596"],"issn-type":[{"value":"1854-3871","type":"electronic"},{"value":"0350-5596","type":"print"}],"subject":[],"published":{"date-parts":[[2026,2,21]]}}}