TL;DR
Federated Test-Time Adaptation (FTTA) allows decentralized models to adapt to local domain shifts during deployment, but it introduces a critical security blind spot. This paper presents FedPoisonTTP, a realistic grey-box data poisoning attack showing that a few compromised clients can severely degrade the performance of honest participants. By generating stealthy, "in-distribution" poisoned inputs that exploit unsupervised test-time objectives, the attack successfully propagates corruption through standard federated aggregation.
What Problem Does This Paper Address?
Machine learning models deployed on edge devices (like smartphones or IoT sensors) frequently encounter "domain shift"—changes in data distribution caused by varying environments, weather, or user behaviors. To combat this without violating user privacy, Federated Test-Time Adaptation (FTTA) has emerged. In FTTA, clients perform lightweight, unsupervised adjustments (like entropy minimization) on their local data streams at test time, periodically sharing updates to refine a global model.
While FTTA improves personalization and robustness, its security risks have been largely ignored. Prior attacks on test-time adaptation assume a "white-box" setting where the attacker has full access to other clients' data or updates. In reality, a malicious federated client only sees the global model broadcasts and participates intermittently. This paper addresses this realistic, highly constrained "grey-box" threat model to see if a constrained adversary can still disrupt the entire network.
Methods
The authors introduce FedPoisonTTP, an attack framework designed to operate under strict stealth and visibility constraints. It relies on three core components:
- Surrogate Aggregator: Since the attacker cannot observe other clients' data or updates, it builds a lightweight surrogate model. It uses history-based extrapolation of past global updates combined with posterior distillation on locally observed responses to predict how the server will aggregate updates.
- Feature-Distribution Regularizer: To ensure the poisoned data transfers effectively to other clients and evades simple anomaly detection, the attack forces the intermediate feature statistics (means and variances) of the poisoned samples to match a local pool of benign data. This keeps the malicious inputs strictly "in-distribution."
- TTA-Aware Objectives: The attack optimizes the poisoned inputs against common test-time adaptation routines using two main strategies:
- Notch High Entropy (NHE): Forces model predictions away from the true class while maximizing output entropy, destabilizing self-training methods like TENT.
- Balanced Low-Entropy (BLE): Nudges the model toward confident but systematically incorrect predictions across classes, avoiding easily detectable class-collapse.
Datasets
The framework was evaluated using two standard computer vision corruption benchmarks:
- CIFAR-10-C
- CIFAR-100-C
Both datasets feature 15 different image corruptions (e.g., blur, noise, brightness changes). The authors focused exclusively on severity level 5 (the worst-case noise) to simulate maximal domain shift and rigorously stress-test the models.
Key Insights and Ideas
- The Vulnerability of Unsupervised Objectives: TTA algorithms rely heavily on unsupervised online objectives like entropy minimization. FedPoisonTTP exploits this by crafting inputs that look completely normal to feature filters but steer the optimization process into harmful local minima.
- The Danger of Small Batch Sizes: Counterintuitively, smaller local batch sizes make clients more vulnerable to the attack. In smaller batches, poisoned samples represent a larger percentage of the adaptation signal, amplifying the disruption before the next server aggregation.
- Stable Transferability: Even though the attacker operates in a grey-box setting with a surrogate model, the relative ranking and effectiveness of the poisoning objectives remain highly stable compared to an ideal white-box attack.
Key Results
The authors tested the attack against popular TTA methods (TENT and CoTTA) across several federated aggregation strategies (FedAvg, FedProx, pFedGraph, FedAMP).
- CIFAR-10-C: Under a standard FedAvg-TENT setup with 5 out of 10 clients compromised (and a 50% local poison ratio), benign client accuracy dropped from 81.19% to 77.95%. For CoTTA, accuracy dropped from 81.61% to 74.82%.
- CIFAR-100-C: The impact was significantly more devastating on the fine-grained CIFAR-100-C dataset. Benign client accuracy under TENT plummeted from 68.13% to 55.15%, while CoTTA dropped from 65.14% to 50.67%—a massive ~14.5% performance gap.
Limitations and Future Work
While the paper successfully exposes a critical security gap in federated test-time personalization, it focuses primarily on demonstrating the feasibility and strength of the attack. The authors do not introduce a novel defensive mechanism to counter FedPoisonTTP, though they show that existing robust aggregation methods (like FedProx or FedAMP) fail to fully mitigate the threat.
Future work will need to focus on developing principled defenses, such as advanced anomaly detection tailored for test-time updates or more resilient collaborative training procedures that can identify feature-consistent but structurally malicious updates.
Summary updated