Attacks¶

Pluggable attacks for adversarial evaluation of sFL-based strategies. Each attack is a subclass of Attack (attacks/base.py) that overrides craft(). Pass the class name to --attack; set --malicious_frac to control the injected fraction of clients.

Available Attacks¶

Name	Description
NoAttack	No-op; benign mode. Packages are returned unmodified (default)
SignFlip	Sends -W instead of W for every malicious client. Destroys gradient signal
GaussianNoise	Replaces malicious updates with zero-mean Gaussian noise scaled to the parameter's own norm
ScaleBoost	Scales honest updates by a constant factor (default 10×) to overpower the aggregate
BackdoorHF	Adds out-of-band high-frequency noise mimicking a spectral trigger; amplitude proportional to parameter norm
StealthHF	Low-amplitude (0.01×) HF noise designed to evade magnitude-based detectors; damage is negligible due to 1/N averaging

Adding a Custom Attack¶

Create a file attacks/MyAttack.py with a class of the same name:

from .base import Attack

class MyAttack(Attack):
    scale: float = 2.0  # optional class-level hyperparameter

    def craft(self, packages, malicious_ids, ctx):
        for cid in malicious_ids:
            pkg = packages[cid]
            pkg["regular_model_params"] = {
                k: v * self.scale
                for k, v in pkg["regular_model_params"].items()
            }
        return packages

ctx is the sFL server instance — use it to access ctx.current_iter, ctx.public_model_params, ctx.configs, etc.

The class is auto-discovered and added to ATTACKS on import; no registration step needed.