攻擊 Surface of Distributed 訓練

"""
Distributed 訓練 topology 安全 analysis.
Models different communication patterns and their 漏洞
to various attack types.
"""
from dataclasses import dataclass, field
from enum import Enum
from typing import Optional
 
 
class TopologyType(Enum):
    PARAMETER_SERVER = "parameter_server"
    ALL_REDUCE_RING = "all_reduce_ring"
    ALL_REDUCE_TREE = "all_reduce_tree"
    GOSSIP = "gossip"
    HIERARCHICAL = "hierarchical"
 
 
@dataclass
class TopologySecurityProfile:
    """安全 profile for a distributed 訓練 topology."""
    topology: TopologyType
    description: str
    num_trust_boundaries: int
    single_point_of_failure: bool
    byzantine_tolerance: int  # Max compromised nodes tolerated
    attack_vectors: list[str]
    detection_capability: str
 
 
TOPOLOGY_PROFILES = [
    TopologySecurityProfile(
        topology=TopologyType.PARAMETER_SERVER,
        description=(
            "Central server aggregates gradients from all workers. "
            "Simple but creates a bottleneck and single point of failure."
        ),
        num_trust_boundaries=2,  # Worker-to-server, server-to-worker
        single_point_of_failure=True,
        byzantine_tolerance=0,  # Server compromise = total compromise
        attack_vectors=[
            "Parameter server compromise (total model control)",
            "Worker gradient 投毒 (amplified by averaging)",
            "Man-in-the-middle on worker-server communication",
            "Denial of service on parameter server",
        ],
        detection_capability="high (centralized 監控)",
    ),
    TopologySecurityProfile(
        topology=TopologyType.ALL_REDUCE_RING,
        description=(
            "Workers form a ring, each sending partial results to the "
            "next. No central server but every worker is on the critical path."
        ),
        num_trust_boundaries=1,  # Worker-to-worker (per link)
        single_point_of_failure=False,
        byzantine_tolerance=0,  # Any worker can corrupt the ring
        attack_vectors=[
            "Gradient corruption propagates through the ring",
            "Worker impersonation in ring topology",
            "Timing attacks exploiting sequential dependencies",
            "Selective corruption of reduce-scatter phase",
        ],
        detection_capability="low (no centralized view)",
    ),
    TopologySecurityProfile(
        topology=TopologyType.ALL_REDUCE_TREE,
        description=(
            "Tree-structured reduction where intermediate nodes aggregate "
            "subtrees. Hierarchical trust model."
        ),
        num_trust_boundaries=2,  # Leaf-to-intermediate, intermediate-to-root
        single_point_of_failure=True,  # Root node
        byzantine_tolerance=0,
        attack_vectors=[
            "Root node compromise affects all gradients",
            "Intermediate node compromise affects its subtree",
            "Leaf worker 投毒 (limited to subtree initially)",
        ],
        detection_capability="medium (hierarchical 監控 possible)",
    ),
    TopologySecurityProfile(
        topology=TopologyType.HIERARCHICAL,
        description=(
            "Multi-level aggregation: intra-node all-reduce, then "
            "inter-node all-reduce. Common in multi-node GPU clusters."
        ),
        num_trust_boundaries=3,  # Intra-node, inter-node, cross-cluster
        single_point_of_failure=False,
        byzantine_tolerance=0,
        attack_vectors=[
            "Intra-node GPU compromise via driver 漏洞",
            "Inter-node communication interception (RDMA/NCCL)",
            "Cross-cluster aggregation manipulation",
            "NVLink/NVSwitch firmware attacks",
        ],
        detection_capability="medium (per-level 監控)",
    ),
]
 
 
def rank_topologies_by_security(
    profiles: list[TopologySecurityProfile],
) -> list[tuple[str, float]]:
    """
    Rank topologies by overall 安全 posture.
    Higher score = more secure.
    """
    scores = []
    for p in profiles:
        score = 0.0
        # Fewer trust boundaries = simpler to secure
        score += max(0, 5 - p.num_trust_boundaries) * 0.2
        # No single point of failure
        score += 0.3 if not p.single_point_of_failure else 0.0
        # Byzantine tolerance
        score += min(0.3, p.byzantine_tolerance * 0.1)
        # 偵測 capability
        detection_scores = {"high": 0.2, "medium": 0.1, "low": 0.0}
        score += detection_scores.get(p.detection_capability, 0.0)
 
        scores.append((p.topology.value, score))
 
    return sorted(scores, key=lambda x: x[1], reverse=True)
 
 
rankings = rank_topologies_by_security(TOPOLOGY_PROFILES)
for topology, score in rankings:
    print(f"  {topology}: 安全 score = {score:.2f}")

"""
Byzantine gradient attack implementations.
Demonstrates various attack strategies available to a
compromised worker node in distributed 訓練.
"""
import numpy as np
from dataclasses import dataclass
from typing import Optional
 
 
@dataclass
class ByzantineAttack:
    """Configuration for a Byzantine gradient attack."""
    name: str
    num_byzantine: int
    total_workers: int
    strategy: str
    stealth_level: str  # "none", "low", "medium", "high"
 
 
def attack_scaling(
    honest_gradients: np.ndarray,
    scale_factor: float = 100.0,
) -> np.ndarray:
    """
    Scaling attack: send gradients that are much larger than normal.
 
    The simplest Byzantine attack — just amplify the gradient to
    dominate the average. Effective against naive averaging but
    easily detected by norm checking.
    """
    mean_gradient = honest_gradients.mean(axis=0)
    return mean_gradient * (-scale_factor)  # Negative to maximize damage
 
 
def attack_sign_flip(
    honest_gradients: np.ndarray,
) -> np.ndarray:
    """
    Sign-flip attack: send the negative of the true gradient.
 
    More subtle than scaling — the magnitude is normal but the
    direction is exactly wrong. Causes 模型 to move away
    from the optimum.
    """
    return -honest_gradients.mean(axis=0)
 
 
def attack_label_flip_gradient(
    honest_gradients: np.ndarray,
    flip_fraction: float = 1.0,
    seed: int = 42,
) -> np.ndarray:
    """
    Label-flip gradient: simulate the gradient that would result from
    訓練 on data with flipped labels.
 
    More sophisticated than sign-flip 因為 the gradient still
    "looks like" a real gradient — it has the right magnitude and
    structure, just from corrupted data.
    """
    rng = np.random.default_rng(seed)
    mean_grad = honest_gradients.mean(axis=0)
 
    # Simulate label-flip by adding noise in a structured way
    noise = rng.standard_normal(mean_grad.shape) * np.std(honest_gradients, axis=0)
    flipped = -mean_grad * flip_fraction + noise * (1 - flip_fraction)
 
    return flipped
 
 
def attack_inner_product_manipulation(
    honest_gradients: np.ndarray,
    target_direction: np.ndarray,
    magnitude_matching: bool = True,
) -> np.ndarray:
    """
    Inner product manipulation (IPM) attack.
 
    Craft a gradient that maximizes the component in the target
    direction while keeping the magnitude consistent with honest
    gradients (if magnitude_matching is True).
 
    這是 one of the strongest known Byzantine attacks 因為
    it is hard to detect with both norm-based and direction-based
    防禦.
    """
    target_norm = target_direction / (np.linalg.norm(target_direction) + 1e-10)
 
    if magnitude_matching:
        # Match the magnitude of honest gradients
        honest_mean_norm = np.mean([
            np.linalg.norm(g) for g in honest_gradients
        ])
        return target_norm * honest_mean_norm
    else:
        return target_direction
 
 
def simulate_byzantine_aggregation(
    honest_gradients: np.ndarray,
    num_byzantine: int,
    attack_fn: callable,
    aggregation: str = "mean",
) -> dict:
    """
    Simulate a Byzantine attack during gradient aggregation.
 
    Args:
        honest_gradients: Shape (num_honest, param_dim).
        num_byzantine: Number of Byzantine workers.
        attack_fn: Function that generates the Byzantine gradient.
        aggregation: Aggregation method ("mean", "median", "trimmed_mean").
 
    Returns:
        Comparison of attack impact under different aggregation methods.
    """
    num_honest = len(honest_gradients)
    total = num_honest + num_byzantine
 
    # Generate Byzantine gradients
    byzantine_grad = attack_fn(honest_gradients)
    byzantine_gradients = np.tile(byzantine_grad, (num_byzantine, 1))
 
    all_gradients = np.concatenate([honest_gradients, byzantine_gradients], axis=0)
 
    # Compute honest-only aggregate (ground truth)
    honest_aggregate = honest_gradients.mean(axis=0)
    honest_norm = np.linalg.norm(honest_aggregate)
 
    # Aggregation under attack
    if aggregation == "mean":
        attacked_aggregate = all_gradients.mean(axis=0)
    elif aggregation == "median":
        attacked_aggregate = np.median(all_gradients, axis=0)
    elif aggregation == "trimmed_mean":
        # Trim top and bottom 10%
        trim_count = max(1, total // 10)
        sorted_grads = np.sort(all_gradients, axis=0)
        attacked_aggregate = sorted_grads[trim_count:-trim_count].mean(axis=0)
    else:
        attacked_aggregate = all_gradients.mean(axis=0)
 
    # Measure impact
    deviation = np.linalg.norm(attacked_aggregate - honest_aggregate)
    cos_sim = float(
        np.dot(attacked_aggregate.flatten(), honest_aggregate.flatten())
        / (np.linalg.norm(attacked_aggregate) * honest_norm + 1e-10)
    )
 
    return {
        "aggregation": aggregation,
        "num_honest": num_honest,
        "num_byzantine": num_byzantine,
        "byzantine_fraction": num_byzantine / total,
        "deviation_from_honest": float(deviation),
        "relative_deviation": float(deviation / (honest_norm + 1e-10)),
        "cosine_similarity": cos_sim,
        "attack_effective": cos_sim < 0.5,
    }
 
 
# Demonstration
np.random.seed(42)
num_honest, param_dim = 20, 256
honest = np.random.randn(num_honest, param_dim) * 0.01
 
print("Byzantine attack effectiveness (1 攻擊者 out of 21 workers):")
print("=" * 60)
 
for attack_name, attack_fn in [
    ("scaling", lambda g: attack_scaling(g, 100)),
    ("sign_flip", attack_sign_flip),
    ("label_flip", attack_label_flip_gradient),
    ("IPM", lambda g: attack_inner_product_manipulation(
        g, np.random.randn(param_dim)
    )),
]:
    for agg in ["mean", "median", "trimmed_mean"]:
        result = simulate_byzantine_aggregation(
            honest, num_byzantine=1, attack_fn=attack_fn, aggregation=agg,
        )
        print(f"  {attack_name} + {agg}: "
              f"cos_sim={result['cosine_similarity']:.3f}, "
              f"effective={result['attack_effective']}")
    print()

"""
Robust gradient aggregation protocols.
Implements Byzantine-tolerant aggregation methods that defend
against compromised worker nodes.
"""
import numpy as np
from typing import Optional
 
 
def krum_aggregation(
    gradients: np.ndarray,
    num_byzantine: int,
    multi_krum: int = 1,
) -> np.ndarray:
    """
    Krum aggregation (Blanchard et al. 2017).
 
    Selects the gradient that is closest to its nearest neighbors,
    effectively excluding outliers. Tolerates up to f Byzantine
    workers when n >= 2f + 3.
 
    Args:
        gradients: Shape (num_workers, param_dim).
        num_byzantine: Maximum number of Byzantine workers.
        multi_krum: Number of gradients to average (Multi-Krum).
    """
    n = len(gradients)
 
    # Compute pairwise distances
    distances = np.zeros((n, n))
    for i in range(n):
        for j in range(i + 1, n):
            dist = np.linalg.norm(gradients[i] - gradients[j])
            distances[i, j] = dist
            distances[j, i] = dist
 
    # 對每個 gradient, compute the sum of distances to its
    # n - num_byzantine - 2 closest neighbors
    num_closest = n - num_byzantine - 2
    if num_closest < 1:
        num_closest = 1
 
    scores = np.zeros(n)
    for i in range(n):
        sorted_distances = np.sort(distances[i])
        # Exclude self (distance 0) and take closest neighbors
        scores[i] = np.sum(sorted_distances[1:num_closest + 1])
 
    # Select gradient(s) with lowest score
    if multi_krum == 1:
        best_idx = np.argmin(scores)
        return gradients[best_idx]
    else:
        best_indices = np.argsort(scores)[:multi_krum]
        return np.mean(gradients[best_indices], axis=0)
 
 
def trimmed_mean_aggregation(
    gradients: np.ndarray,
    trim_fraction: float = 0.1,
) -> np.ndarray:
    """
    Coordinate-wise trimmed mean aggregation.
 
    對每個 parameter coordinate, sorts the values across workers
    and trims the top and bottom fraction before averaging.
    這是 robust against a bounded number of Byzantine workers.
    """
    n = len(gradients)
    trim_count = max(1, int(n * trim_fraction))
 
    sorted_grads = np.sort(gradients, axis=0)
    trimmed = sorted_grads[trim_count:n - trim_count]
 
    return np.mean(trimmed, axis=0)
 
 
def geometric_median_aggregation(
    gradients: np.ndarray,
    max_iterations: int = 100,
    tolerance: float = 1e-6,
) -> np.ndarray:
    """
    Geometric median aggregation (Weiszfeld's algorithm).
 
    The geometric median minimizes the sum of distances to all
    points, making it robust to outliers. Unlike coordinate-wise
    methods, it considers the full gradient vector geometry.
    """
    estimate = np.mean(gradients, axis=0)
 
    for _ in range(max_iterations):
        distances = np.array([
            np.linalg.norm(g - estimate) for g in gradients
        ])
        # Avoid division by zero
        distances = np.maximum(distances, 1e-10)
 
        weights = 1.0 / distances
        weights /= weights.sum()
 
        new_estimate = np.average(gradients, weights=weights, axis=0)
 
        if np.linalg.norm(new_estimate - estimate) < tolerance:
            break
        estimate = new_estimate
 
    return estimate
 
 
def evaluate_aggregation_robustness(
    honest_gradients: np.ndarray,
    byzantine_gradients: np.ndarray,
    methods: dict[str, callable],
) -> dict:
    """
    評估 how well different aggregation methods resist
    Byzantine attacks.
    """
    all_gradients = np.concatenate(
        [honest_gradients, byzantine_gradients], axis=0
    )
    honest_truth = np.mean(honest_gradients, axis=0)
    honest_norm = np.linalg.norm(honest_truth)
 
    results = {}
    for name, method in methods.items():
        aggregated = method(all_gradients)
        deviation = np.linalg.norm(aggregated - honest_truth)
        cos_sim = float(
            np.dot(aggregated.flatten(), honest_truth.flatten())
            / (np.linalg.norm(aggregated) * honest_norm + 1e-10)
        )
        results[name] = {
            "deviation": float(deviation),
            "relative_deviation": float(deviation / (honest_norm + 1e-10)),
            "cosine_similarity": cos_sim,
        }
 
    return results
 
 
# Compare robust aggregation methods
np.random.seed(42)
n_honest, n_byzantine, dim = 15, 5, 128
honest = np.random.randn(n_honest, dim) * 0.01
# Strong scaling attack
byzantine = np.tile(-honest.mean(axis=0) * 50, (n_byzantine, 1))
 
methods = {
    "mean": lambda g: np.mean(g, axis=0),
    "krum": lambda g: krum_aggregation(g, n_byzantine),
    "trimmed_mean": lambda g: trimmed_mean_aggregation(g, 0.2),
    "geometric_median": lambda g: geometric_median_aggregation(g),
}
 
results = evaluate_aggregation_robustness(honest, byzantine, methods)
print(f"Scenario: {n_byzantine}/{n_honest + n_byzantine} workers compromised")
print(f"攻擊: scaling (50x)")
print()
for name, r in results.items():
    status = "DEFENDED" if r["cosine_similarity"] > 0.8 else "COMPROMISED"
    print(f"  {name:20s}: cos_sim={r['cosine_similarity']:.3f} [{status}]")

"""
Communication channel 安全 評估 for distributed 訓練.
Analyzes the 安全 properties of different communication backends.
"""
from dataclasses import dataclass, field
 
 
@dataclass
class CommChannelProfile:
    """安全 profile for a communication channel."""
    name: str
    protocol: str
    encryption_support: bool
    authentication_support: bool
    integrity_verification: bool
    performance_overhead_with_security: str  # "low", "medium", "high"
    known_vulnerabilities: list[str] = field(default_factory=list)
 
 
COMM_PROFILES = [
    CommChannelProfile(
        name="NCCL over TCP",
        protocol="TCP/IP",
        encryption_support=False,
        authentication_support=False,
        integrity_verification=False,
        performance_overhead_with_security="high",
        known_vulnerabilities=[
            "No encryption — gradients visible to network observers",
            "No 認證 — worker impersonation possible",
            "No integrity checking — gradient modification in transit",
        ],
    ),
    CommChannelProfile(
        name="NCCL over InfiniBand/RDMA",
        protocol="RDMA",
        encryption_support=False,
        authentication_support=True,  # Fabric-level
        integrity_verification=False,
        performance_overhead_with_security="medium",
        known_vulnerabilities=[
            "RDMA bypasses OS network stack (no firewall)",
            "Memory registration exposes device memory",
            "Limited encryption support at wire speed",
            "Physical access to fabric = full access",
        ],
    ),
    CommChannelProfile(
        name="Gloo (PyTorch)",
        protocol="TCP/IP",
        encryption_support=True,  # TLS possible
        authentication_support=True,
        integrity_verification=True,  # With TLS
        performance_overhead_with_security="high",
        known_vulnerabilities=[
            "TLS not enabled by default",
            "Certificate management complexity",
            "Significant performance penalty with encryption",
        ],
    ),
]
 
 
def assess_communication_security(
    profiles: list[CommChannelProfile],
    requires_encryption: bool = True,
    requires_authentication: bool = True,
) -> list[dict]:
    """評估 each communication channel against 安全 requirements."""
    assessments = []
    for profile in profiles:
        gaps = []
        if requires_encryption and not profile.encryption_support:
            gaps.append("NO_ENCRYPTION")
        if requires_authentication and not profile.authentication_support:
            gaps.append("NO_AUTHENTICATION")
        if not profile.integrity_verification:
            gaps.append("NO_INTEGRITY_CHECK")
 
        assessments.append({
            "channel": profile.name,
            "security_gaps": gaps,
            "compliant": len(gaps) == 0,
            "risk_level": (
                "critical" if len(gaps) >= 2
                else "high" if len(gaps) == 1
                else "acceptable"
            ),
        })
 
    return assessments
 
 
assessments = assess_communication_security(COMM_PROFILES)
for a in assessments:
    print(f"{a['channel']}: {a['risk_level']} "
          f"(gaps: {', '.join(a['security_gaps']) or 'none'})")