Implement adaptive attacks: stat-opt, dny-opt, min-max, min-sum

[Copilot]

Adds four adaptive attack strategies that learn from defense responses and optimize evasion. Each attack implements a distinct optimization approach based on published adversarial FL research.

Implementations

Base Infrastructure (adaptive_base.py)

• AdaptiveAttack class with feedback collection, detection/acceptance rate tracking, and adaptation hooks
• All adaptive attacks inherit common feedback mechanism

Attack Strategies

• stat-opt: Constrains malicious updates within statistical bounds (mean ± k·σ) of benign clients. Adapts constraint factor based on detection rate.
• dny-opt: Q-learning with ε-greedy exploration across attack intensities and techniques (sign flip, gradient noise, scaling). Updates Q-values based on acceptance/rejection rewards.
• min-max: Game-theoretic optimization assuming optimal defender. Evaluates attack performance across defense ensemble (Krum, trimmed mean, median). Adapts threat model weights.
• min-sum: Minimizes sum of distances to benign updates via gradient descent optimization. Balances stealth (distance minimization) vs impact (attack magnitude).

Integration

Client Runner (client_runner.py)
Extended create_attack() to instantiate adaptive attacks from YAML configs with attack-specific parameters.

Configuration (config.py)
AttackConfig dataclass now includes parameters for all four attacks (constraint_factor, learning_rate, defense_models, distance_weight, etc.). Default initialization ensures uniform threat model weights.

Usage

from src.attacks import StatOptAttack, DnyOptAttack, MinMaxAttack, MinSumAttack

# Statistical optimization - evades statistical defenses
stat_attack = StatOptAttack(
    intensity=0.2,
    constraint_factor=1.5,  # Within 1.5σ of benign mean
    target_clients=[0, 1, 2]
)

# Q-learning adaptation
dny_attack = DnyOptAttack(
    intensity=0.15,
    learning_rate=0.1,
    exploration_rate=0.1
)

# Game-theoretic with defense ensemble
minmax_attack = MinMaxAttack(
    intensity=0.2,
    defense_models=['krum', 'trimmed_mean', 'cognitive']
)

# Distance minimization
minsum_attack = MinSumAttack(
    intensity=0.2,
    distance_weight=0.7  # 70% weight on distance minimization
)

YAML Configuration

attacks:
  - enabled: true
    attack_type: "stat_opt"
    intensity: 0.2
    constraint_factor: 1.5
    target_clients: [0, 1, 2]

Documentation

• docs/ADAPTIVE_ATTACKS.md: Algorithm descriptions, mathematical formulations, defense evasion strategies, academic references (Fang et al. USENIX'20, Baruch et al. NeurIPS'19, Shejwalkar & Houmansadr NDSS'21, Bhagoji et al. ICML'19)
• docs/IMPLEMENTATION_SUMMARY.md: Implementation details, architecture, integration points
• Example configs: experiments/configs/{stat_opt,dny_opt,min_max,min_sum}_attack_test.yaml

Files Changed

• New: 11 files (4 attack implementations, base class, 5 configs, 2 docs)
• Modified: 5 files (client_runner, config, init, README, test_local_setup)
• Total: +2,209 lines

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.