LLM 的有害基因：大模型生成有害内容的统一内在机制被发现

import torch
import torch.nn as nn
from typing import Dict, List, Tuple

class WeightImportanceScorer:
    """计算 LLM 各层权重对有害内容生成的重要性评分"""

    def __init__(self, model: nn.Module):
        self.model = model
        self.harmful_scores: Dict[str, float] = {}

    def evaluate_harmfulness(self, prompts: List[str]) -> float:
        """评估模型在一组有害提示上的有害性得分"""
        total_score = 0.0
        with torch.no_grad():
            for prompt in prompts:
                outputs = self.model.generate(prompt, max_length=50)
                score = self._classify_harmfulness(outputs)
                total_score += score
        return total_score / len(prompts)

    def _classify_harmfulness(self, text: str) -> float:
        """简化的有害性分类器（返回 0-1 之间的分数）"""
        harmful_keywords = ["暴力", "仇恨", "歧视", "伤害"]
        matches = sum(1 for kw in harmful_keywords if kw in text)
        return min(matches / 3.0, 1.0)

    def score_layer_importance(
        self, layer_name: str, harmful_prompts: List[str]
    ) -> float:
        """计算单个层对有害内容生成的重要性"""
        # 基准有害性得分
        baseline = self.evaluate_harmfulness(harmful_prompts)

        # 临时移除该层后重新评估
        original_weights = self._zero_out_layer(layer_name)
        perturbed = self.evaluate_harmfulness(harmful_prompts)
        self._restore_layer(layer_name, original_weights)

        # 重要性 = 基准得分与扰动后得分的差异
        importance = baseline - perturbed
        self.harmful_scores[layer_name] = importance
        return importance

    def _zero_out_layer(self, layer_name: str) -> torch.Tensor:
        """临时将某层权重置零并返回原始值"""
        layer = dict(self.model.named_parameters())[layer_name]
        original = layer.data.clone()
        layer.data.zero_()
        return original

    def _restore_layer(self, layer_name: str, weights: torch.Tensor):
        """恢复某层的原始权重"""
        layer = dict(self.model.named_parameters())[layer_name]
        layer.data.copy_(weights)


# === 使用示例 ===
# scorer = WeightImportanceScorer(my_llm)
# importance = scorer.score_layer_importance("layers.12.weight", harmful_prompts)
# print(f"Layer 12 importance: {importance:.4f}")

from dataclasses import dataclass
from typing import Dict, List
import numpy as np


@dataclass
class PruningResult:
    """剪枝实验结果"""
    category: str
    harmful_rate_before: float  # 剪枝前有害内容生成率
    harmful_rate_after: float   # 剪枝后有害内容生成率
    benign_score_before: float  # 剪枝前正常能力得分
    benign_score_after: float   # 剪枝后正常能力得分


def simulate_harmful_pruning(
    model_name: str,
    harmful_categories: List[str],
    prune_ratio: float = 0.15,
) -> List[PruningResult]:
    """
    模拟有害权重剪枝实验
    基于 arXiv:2604.09544 的发现
    """
    results = []

    # 模拟：有害权重子集跨所有类型通用
    unified_harmful_weights = _identify_unified_weights(model_name)

    for category in harmful_categories:
        # 剪枝前评估
        before_harm = _test_harmful_generation(
            model_name, category, unified_harmful_weights
        )
        before_benign = _test_benign_capabilities(model_name)

        # 执行剪枝
        _prune_weights(model_name, unified_harmful_weights, prune_ratio)

        # 剪枝后评估
        after_harm = _test_harmful_generation(
            model_name, category, unified_harmful_weights
        )
        after_benign = _test_benign_capabilities(model_name)

        results.append(PruningResult(
            category=category,
            harmful_rate_before=before_harm,
            harmful_rate_after=after_harm,
            benign_score_before=before_benign,
            benign_score_after=after_benign,
        ))

    return results


def _identify_unified_weights(model: str) -> np.ndarray:
    """识别跨有害类型通用的权重子集"""
    # 简化模拟：返回一个掩码数组
    np.random.seed(42)
    return np.random.random(1000) < 0.15  # 15% 的权重被标记为有害


def _test_harmful_generation(model, category, weights) -> float:
    """测试某类有害内容的生成率"""
    return np.random.uniform(0.3, 0.6)  # 模拟值


def _test_benign_capabilities(model) -> float:
    """测试正常能力得分"""
    return np.random.uniform(0.85, 0.95)


def _prune_weights(model, weights, ratio):
    """执行权重剪枝"""
    pass  # 实际操作需要修改模型参数


# === 运行实验并打印结果 ===
categories = ["暴力内容", "仇恨言论", "虚假信息", "恶意建议"]
results = simulate_harmful_pruning("LLaMA-3-70B", categories)

print(f"{'类别':<12} | {'剪枝前有害%':>10} | {'剪枝后有害%':>10} | {'正常能力变化':>10}")
print("-" * 52)
for r in results:
    harm_reduction = ((r.harmful_rate_before - r.harmful_rate_after) / r.harmful_rate_before * 100)
    benign_change = r.benign_score_after - r.benign_score_before
    print(f"{r.category:<10} | {r.harmful_rate_before:>9.1%} | {r.harmful_rate_after:>9.1%} | {benign_change:>+9.3f}")