SignSGD with Federated Defense: Harnessing Adversarial Attacks through Gradient Sign Decoding

Chanho Park, Namyoon Lee

TL;DR

This paper introduces signSGD-FD, a novel distributed learning algorithm that maintains fast convergence even when a large portion of the network is malicious. By treating adversarial workers as "noisy communication channels" and using coding theory to decode their true intent, the authors show that training performance can remain virtually unchanged as long as honest workers outnumber the attackers.

What Problem Does This Paper Address?

Training massive AI models requires distributed learning, where many "workers" compute gradients in parallel. However, moving full-precision gradients between workers and a central server creates a massive communication bottleneck. SignSGD solves this by sending only the sign of the gradient (1 bit per parameter), drastically reducing bandwidth.

The catch? SignSGD is notoriously fragile. If a few "Byzantine" (adversarial) workers deliberately flip their gradient signs or provide junk data, the standard "Majority Voting" (MV) aggregation fails, and the model's accuracy plummets. Existing defenses usually try to identify and "filter out" these bad actors, which is computationally expensive and often throws away potentially useful information.

Methods

The authors propose signSGD with Federated Defense (signSGD-FD). Instead of simple majority voting, they use a coding-theoretical approach called Weighted Majority Voting (WMV).

  1. Communication as a Channel: They model each worker as a Binary Symmetric Channel (BSC). An honest worker has a low probability of sending the wrong sign, while an adversary has a high probability (or even a 100% probability in a "Sign Inversion Attack").
  2. Reliability Estimation: During an initial "warm-up" phase, the server tracks how often each worker's submitted sign disagrees with the consensus. It uses this to estimate a "cross-over probability" for every worker.
  3. Log-Likelihood Ratio (LLR) Weights: The server assigns a weight to each worker based on their estimated reliability.
  4. Highly reliable workers get high positive weights.
  5. Workers who are pure noise ($r=0.5$) get weights near zero.
  6. The "Aha!" Moment: Workers who consistently flip their signs (attackers) are assigned negative weights. By multiplying a flipped sign by a negative weight, the server effectively "decodes" the attacker's information and uses it to help the model converge.

Key Insights and Ideas

  • Adversaries are Informative: Traditional wisdom says attackers destroy information. This paper argues that if an attacker’s behavior is consistent (even if it's consistently malicious), they are actually providing information that can be "decoded" and used for training.
  • The Invariance Principle: Theoretically, the convergence rate of signSGD-FD remains invariant regardless of the number of attackers, provided that the number of honest workers is greater than the number of adversarial workers ($M-L > 0$).
  • The Worst Attack is Randomness: Counter-intuitively, the most damaging attack isn't a worker who always lies ($r=1$); it's a worker who flips their sign exactly 50% of the time ($r=0.5$). At 50%, the worker becomes a source of pure entropy, providing zero useful information to the server.
Test accuracy vs. sign-flipping probability. signSGD-FD maintains high accuracy even when adversaries consistently flip signs (r=1.0), whereas standard signSGD-MV fails.

Key Results

The researchers evaluated their approach using standard image classification benchmarks (MNIST, CIFAR-10, CIFAR-100) against various attack scenarios, including Sign Inversion Attacks (SIA) and Stochastic Sign Flip Attacks (SSFA).

  • Robustness: In tests with a Sign Inversion Attack, standard signSGD-MV failed to converge when 9 out of 15 workers were attackers. In contrast, signSGD-FD maintained nearly the same accuracy as a system with zero attackers.
  • Communication Efficiency: Compared to other robust algorithms like Multi-Krum, signSGD-FD reduced communication costs by 30x while achieving higher test accuracy.
  • Superiority over Baselines: signSGD-FD outperformed other specialized robust optimizers (like Sto-signSGD and Election-signSGD) across all tested datasets, especially in high-intensity attack scenarios.
Test accuracy over communication rounds. signSGD-FD (solid lines) remains robust even with 9 adversaries (L=9), while signSGD-MV (dashed lines) completely fails to converge.

Limitations and Future Work

  • The Majority Constraint: The current proof and algorithm rely on honest workers outnumbering adversarial ones. If attackers control more than 50% of the nodes, the initial reliability estimation may fail.
  • Black-Box Assumption: The paper assumes a "black-box" setting where attackers cannot see the data of honest workers. If attackers could coordinate based on the global state in real-time, the defense might need further hardening.
  • Initial Phase Sensitivity: The performance depends on the "initial phase" ($T_{in}$) being long enough to accurately estimate worker weights. Finding the optimal $T_{in}$ automatically for different model architectures remains an open area for improvement.

Summary updated

Abstract

Distributed learning is an effective approach to accelerate model training by using parallel computing power of multiple workers. However, substantial communication delays arise between workers and a parameter server due to the massive costs associated with communicating gradients. SignSGD with majority voting (signSGD-MV) is a simple yet effective optimizer that reduces communication costs through sign quantization, but its convergence rate significantly decreases when adversarial workers arbitrarily manipulate datasets or local gradient updates. In this paper, we consider a distributed learning problem where the workforce comprises a mixture of honest and adversarial workers. In this setting, we show that the convergence rate can remain invariant as long as the number of honest workers providing trustworthy local updates to the parameter server exceeds the number of adversarial workers. The key idea behind this counter-intuitive result is our novel aggregation method, signSGD with federated defense (signSGD-FD). Unlike traditional approaches, signSGD-FD utilizes the gradient information sent by adversarial workers with appropriate weights, obtained through gradient sign decoding. Experimental results demonstrate that signSGD-FD achieves superior convergence rates compared to traditional algorithms in various adversarial attack scenarios.
BibTeX
@inproceedings{park2024signsgd,
  author = {Chanho Park and Namyoon Lee},
  title = {{SignSGD with Federated Defense: Harnessing Adversarial Attacks through Gradient Sign Decoding}},
  booktitle = {International Conference on Machine Learning},
  year = {2024},
  url = {https://openreview.net/forum?id=zEqeNEuiJr},
}
Figure