Title: "The Challenges of Governance in Federated Learning," 

Abstract: The Internet provides nearly 50 years of lessons in federated systems, starting from E-Mail, from the 1970s, the Domain system in the 1980s, Inter-domain Routing from the 1990s, and the "Fediverse" in the 2020s. Each of these systems has seen evolution over the years. For example, recently E-Mail systems adopted DMARC and DKIM, to reduce the prevalance of spam from unaccountable sources. Certificates and CA Transparency (decentralised)  was added to the DNS.
BGP speakers have to check a (somewhat) centralised database to infer whether an AS owns an IP. And the fediverse, epitomised by BlueSky and Mastodon, has decentralised rules for content moderation.

The case for federated learning is potentially much harder. In this talk, I will outline some of the challenges, starting by example from the cases above, but moving on to discuss threats (model drift, se membership inference and model inversion attacks and more). Despite that federated learning is inherently cooperative, it suffers from all the potential risks (denial & theft of service, pollution/poisoning attacks and so on) that the Internet layers have witnessed in the past. Because it is also a general purpose technology, it is not amenable to simple, static one-shot solution designs, and research is needed to even understand what might be viable ways to provide goveranance (future stability) without simply re-centralising.

Speaker: Jon Crowcroft, 
Bio: https://en.wikipedia.org/wiki/Jon_Crowcroft
Time&Place: 8:00 BST, 24th Oct 2025, online, zoom or teams, tba.