In Federated Learning (FL), a central server aggregates model updates from thousands of decentralized clients. While this preserves data privacy, it creates a unique security vulnerability: a single malicious participant can inject a "backdoor" trigger into the global model.
Historically, these attacks were considered fragile. Because the global model is constantly being updated by a "continuous stream of benign updates," the malicious modifications are typically "washed away" or overwritten within a few rounds—a phenomenon akin to catastrophic forgetting. Recent research, however, has shifted toward making these backdoors durable, ensuring they persist long after the attacker has left the training loop.
The Erasure Problem: Why Backdoors Fade
In a standard FL round, the server averages updates from a random subset of clients. If one attacker injects a backdoor, their specific weight modifications are immediately diluted by the honest updates of the other participants in that round. Even if the attack succeeds initially, subsequent rounds of training on clean data act as a natural defense, gradually "unlearning" the adversarial behavior.
As noted by Ashwinee Panda and colleagues in Neurotoxin: Durable Backdoors in Federated Learning, this lack of durability is the primary bottleneck for real-world exploits. If a backdoor only lasts for five rounds in a system that trains for thousands, the attacker has almost no chance of the trigger being active when the model is actually deployed.
Strategy 1: Targeting the "Quiet" Parameters
The most intuitive way to ensure a backdoor survives is to hide it where benign updates rarely go. In any neural network, gradients are sparse; in any given training round, only a fraction of the model's parameters receive significant updates.
The Neurotoxin approach, developed by Ashwinee Panda and Zhengming Zhang, involves identifying these "infrequently updated" coordinates. By restricting malicious updates to parameters that honest clients aren't using, the attacker ensures their changes aren't overwritten. Thuy Dung Nguyen and Tuan Minh Nguyen furthered this in IBA: Towards Irreversible Backdoor Attacks in Federated Learning, demonstrating that targeting these "quiet" parameters makes the backdoor nearly impossible to erase through standard fine-tuning.
Similarly, Ki Yoon Yoo showed in Backdoor Attacks in Federated Learning by Rare Embeddings and Gradient Ensembling that NLP models are particularly vulnerable because rare word embeddings are almost never updated by benign users, providing a permanent "parking spot" for malicious triggers.
Strategy 2: Anchoring in the Feature Space
Rather than just hiding in the weights, an attacker can try to make the backdoor "look" like legitimate data. If the backdoor trigger is mathematically similar to the features of the target class, benign updates to that class might actually reinforce the backdoor instead of erasing it.
Yanbo Dai and Songze Li introduced this concept with Chameleon (Chameleon: Adapting to Peer Images for Planting Durable Backdoors in Federated Learning). They used contrastive learning to align poisoned samples with "facilitator" images from the target class. By anchoring the backdoor in the model's primary activation paths, they extended the attack's lifespan by up to 4x.
A similar intuition drives Mirage, proposed by Ye Li and Yanchao Zhao in Infighting in the Dark: Multi-Label Backdoor Attack in Federated Learning. By hiding triggers within the normal data distribution, they ensure the malicious behavior is deeply integrated into the model's logic rather than existing as a fragile, peripheral weight modification.
Strategy 3: Anticipatory Optimization
A third direction involves treating the FL aggregation process as a known obstacle to be optimized against. Instead of crafting an update that works now, attackers can craft an update that is designed to survive the next several rounds of averaging.
In Thinking Two Moves Ahead: Anticipating Other Users Improves Backdoor Attacks in Federated Learning, Yuxin Wen and Jonas Geiping introduced the Anticipate attack. By simulating future training rounds during the local optimization phase, the attacker can steer the model toward a region of the loss landscape that is resistant to the "averaging out" effect.
Similarly, Hangfan Zhang and Jinyuan Jia developed A3FL (A3FL: Adversarially Adaptive Backdoor Attacks to Federated Learning), which uses adversarial adaptation to ensure triggers remain effective even as the global model evolves.
The Defender's Challenge: Backdoor Leakage
The shift toward durable attacks has revealed a major flaw in current defenses. Most Byzantine-resilient aggregation methods (like Krum or Median) focus on filtering out "outlier" updates in a single round.
However, Shuaiqi Wang and Giulia Fanti identify a phenomenon called backdoor leakage in Towards a Defense against Backdoor Attacks in Continual Federated Learning. They argue that even if a defense is 99% effective at filtering malicious updates, the remaining 1% of "residual signals" can accumulate over thousands of rounds. In a continual learning setting, these tiny, persistent errors eventually consolidate into a successful backdoor, making one-shot defense evaluations insufficient for real-world security.