My federated learning project just showed that "high accuracy" can completely hide a model missing every single attack from an entire category, and I think more people should know about this [R]

Setup

I've been working on a research project comparing federated learning algorithms (FedAvg, FedProx, FedNova) against a centralized baseline for network intrusion detection, using the CICIDS2017 dataset split across four simulated "silos" by attack type. Three of the silos have tons of data, but one silo (Web Attacks) only has about 3k samples out of 3 million total, so it's a pretty extreme imbalance.

The Problem: Global Accuracy Hides Failure

The thing that got me was how good the global accuracy numbers looked while completely hiding what was actually happening underneath. FedAvg was hitting like 96% global accuracy, which sounds great, but when I broke it down by silo, the minority silo was sitting at like 49% accuracy with literally 0.00 recall on the attack class, meaning it missed every single attack in that category. The global number just averages that out because the big silos are doing fine and there's so much more data in them, so the failure basically gets buried.

Centralized Baseline Instability

Even weirder, I reran the centralized model (the "gold standard" baseline that gets to see all the data at once, no federation at all) across 10 different random seeds just to sanity check things, and its performance on that same minority silo swung from about 57% to 99.5% depending purely on the seed. Same model, same data, same everything except the random seed, and it either completely nails the rare attack class or completely whiffs on it. That kind of instability in a "centralized is the safe baseline" model was not what I expected going in.

FedNova Consistency

FedNova (which normalizes updates by local step count instead of just averaging) ended up being way more consistent across all silos, staying in the high 90s no matter which silo or seed, without giving up any global accuracy either.

Conclusion

So the actual conclusion of the paper is basically: global accuracy is not a trustworthy metric on its own in federated intrusion detection, you have to look at per-client performance, and picking your aggregation method actually matters a lot more for rare attack detection than the global number would ever suggest. Currently rewriting this for a conference submission and happy to answer questions if anyone's curious about the setup or findings.