# 导入必要的库
import torch
import numpy as np
import matplotlib.pyplot as plt
from transformers import AutoModelForSequenceClassification, AutoTokenizer

# 设置中文显示
plt.rcParams['font.sans-serif'] = ['SimHei', 'DejaVu Sans']
plt.rcParams['axes.unicode_minus'] = False

print("正在加载中文情感分析模型...")

# 使用中文情感分析模型
model_name = "uer/roberta-base-finetuned-jd-binary-chinese"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
model.eval()

print("✓ 模型加载完成！")

# 定义中文同义词词典（简化版）
# 为了教学目的，我们手动定义常见词的同义词
SYNONYMS = {
    # 正面词汇
    "好": ["棒", "佳", "优", "赞"],
    "喜欢": ["爱", "青睐", "偏爱", "钟爱"],
    "优秀": ["出色", "杰出", "卓越", "优良"],
    "满意": ["称心", "如意", "舒心", "惬意"],
    "推荐": ["建议", "介绍", "举荐", "力荐"],
    "快": ["迅速", "敏捷", "飞快", "神速"],
    "便宜": ["实惠", "划算", "低价", "优惠"],
    "漂亮": ["美丽", "好看", "靓丽", "精美"],
    
    # 负面词汇
    "差": ["糟", "烂", "劣", "次"],
    "讨厌": ["厌恶", "反感", "嫌弃", "憎恨"],
    "失望": ["沮丧", "灰心", "气馁", "心寒"],
    "慢": ["缓慢", "迟缓", "磨蹭", "拖拉"],
    "贵": ["昂贵", "高价", "天价", "奢侈"],
    "难用": ["难操作", "不好用", "复杂", "麻烦"],
    
    # 中性词汇
    "产品": ["商品", "货物", "物品", "东西"],
    "购买": ["买", "选购", "入手", "下单"],
    "使用": ["用", "运用", "利用", "采用"],
}

def get_synonyms(word):
    """获取词的同义词列表"""
    return SYNONYMS.get(word, [])

print(f"同义词词典包含 {len(SYNONYMS)} 个词条")
print(f"示例：'喜欢' 的同义词 = {get_synonyms('喜欢')}")

# 辅助函数
def analyze_sentiment(text):
    """分析文本情感，返回标签和置信度"""
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
    with torch.no_grad():
        outputs = model(**inputs)
        probs = torch.softmax(outputs.logits, dim=1)[0]
    
    # 假设 label 0 = 负面, label 1 = 正面
    if probs[1] > probs[0]:
        return "正面", probs[1].item()
    else:
        return "负面", probs[0].item()

def get_positive_score(text):
    """获取正面情感的置信度"""
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
    with torch.no_grad():
        outputs = model(**inputs)
        probs = torch.softmax(outputs.logits, dim=1)[0]
    return probs[1].item()

# 测试
test_texts = [
    "这个产品非常好，我很喜欢！",
    "太差了，完全不推荐购买。",
    "质量一般，价格还行。"
]

print("情感分析测试：")
for text in test_texts:
    label, score = analyze_sentiment(text)
    print(f"  '{text}' → {label} ({score:.2%})")

# 【填空 1】实现词重要性评估函数
# 提示：通过删除每个词，观察模型输出的变化来评估重要性

def compute_word_importance(text):
    """
    计算每个词对预测结果的重要性
    方法：删除该词后，观察置信度变化
    """
    # 简单分词（按字符）
    chars = list(text)
    original_score = get_positive_score(text)
    importance_scores = []
    
    for i, char in enumerate(chars):
        if char in ['，', '。', '！', '？', ' ']:  # 跳过标点
            importance_scores.append((char, 0))
            continue
            
        # 创建删除第 i 个字后的文本
        chars_without_i = chars[:i] + chars[i+1:]
        modified_text = ''.join(chars_without_i)
        
        if modified_text.strip():  # 确保不是空文本
            # 【填空 1】计算删除该字后的情感分数
            # 参考答案：modified_score = get_positive_score(modified_text)
            modified_score = ___________________
            
            # 重要性 = 删除后分数变化的绝对值
            importance = abs(original_score - modified_score)
        else:
            importance = 0
        
        importance_scores.append((char, importance))
    
    return importance_scores

# 测试词重要性
test_sentence = "这个产品非常好，我很喜欢"
importance = compute_word_importance(test_sentence)

print(f"原句：'{test_sentence}'")
print(f"原始情感：{analyze_sentiment(test_sentence)}")
print("\n字重要性排序（前10）：")
for char, score in sorted(importance, key=lambda x: x[1], reverse=True)[:10]:
    if score > 0:
        print(f"  '{char}': {score:.4f}")

# 可视化词重要性
chars = [c for c, _ in importance]
scores = [s for _, s in importance]

plt.figure(figsize=(14, 4))
colors = ['red' if s > 0.05 else 'steelblue' for s in scores]
plt.bar(range(len(chars)), scores, color=colors)
plt.xticks(range(len(chars)), chars, fontsize=12)
plt.xlabel('字符')
plt.ylabel('重要性分数')
plt.title(f'字符重要性分析\n"{test_sentence}"')
plt.tight_layout()
plt.show()

print("红色标记的字对情感判断影响最大，是攻击的优先目标")

# 【填空 2】实现同义词替换攻击
# 提示：找到文本中可替换的词，尝试替换后观察效果

def synonym_attack(text):
    """
    同义词替换攻击
    
    参数：
        text: 原始文本
    
    返回：
        对抗文本, 是否成功, 替换历史
    """
    original_label, original_conf = analyze_sentiment(text)
    current_text = text
    replacements = []
    
    # 遍历同义词词典中的每个词
    for word, syns in SYNONYMS.items():
        if word not in current_text:
            continue
        
        # 尝试每个同义词
        best_synonym = None
        best_score_change = 0
        
        for synonym in syns:
            # 【填空 2】创建替换后的文本
            # 参考答案：test_text = current_text.replace(word, synonym, 1)
            test_text = ___________________
            
            new_label, new_conf = analyze_sentiment(test_text)
            
            # 如果标签改变了，攻击成功
            if new_label != original_label:
                replacements.append((word, synonym))
                return test_text, True, replacements
            
            # 记录使置信度下降最多的替换
            if original_label == "正面":
                score_change = original_conf - new_conf
            else:
                new_pos_score = get_positive_score(test_text)
                orig_pos_score = get_positive_score(current_text)
                score_change = new_pos_score - orig_pos_score
            
            if score_change > best_score_change:
                best_score_change = score_change
                best_synonym = synonym
        
        # 应用最佳替换
        if best_synonym and best_score_change > 0.01:
            replacements.append((word, best_synonym))
            current_text = current_text.replace(word, best_synonym, 1)
            
            # 检查是否攻击成功
            new_label, _ = analyze_sentiment(current_text)
            if new_label != original_label:
                return current_text, True, replacements
    
    return current_text, False, replacements

# 测试同义词攻击
test_sentences = [
    "这个产品非常好，我很喜欢！",
    "质量优秀，价格便宜，强烈推荐！",
    "发货快，产品漂亮，很满意。",
    "太差了，完全不推荐购买。",
]

print("=" * 60)
print("同义词替换攻击测试")
print("=" * 60)

for sentence in test_sentences:
    orig_label, orig_conf = analyze_sentiment(sentence)
    adv_text, success, replacements = synonym_attack(sentence)
    adv_label, adv_conf = analyze_sentiment(adv_text)
    
    print(f"\n原句: {sentence}")
    print(f"原始: {orig_label} ({orig_conf:.2%})")
    print(f"替换: {replacements}")
    print(f"对抗: {adv_text}")
    print(f"结果: {adv_label} ({adv_conf:.2%}) - {'攻击成功 ✓' if success else '攻击失败 ✗'}")

# 【填空 3】批量测试攻击成功率
# 提示：在多个样本上测试，统计成功率

test_samples = [
    "这款手机非常好用，我很喜欢。",
    "产品质量优秀，物流也快。",
    "服务态度好，产品漂亮。",
    "价格便宜，性价比高，推荐购买。",
    "太差了，质量很差。",
    "发货慢，产品难用。",
    "很失望，不推荐。",
    "贵而且质量差。",
]

success_count = 0
total_replacements = 0

print("批量攻击测试：")
print("-" * 60)

for sample in test_samples:
    # 【填空 3】执行攻击并统计结果
    # 参考答案：adv_text, success, replacements = synonym_attack(sample)
    adv_text, success, replacements = ___________________
    
    if success:
        success_count += 1
        total_replacements += len(replacements)
        print(f"✓ {sample[:25]}... -> 替换 {len(replacements)} 个词")
    else:
        print(f"✗ {sample[:25]}...")

print("-" * 60)
print(f"攻击成功率：{success_count}/{len(test_samples)} ({success_count/len(test_samples)*100:.1f}%)")
if success_count > 0:
    print(f"平均替换词数：{total_replacements/success_count:.1f}")

# 可视化攻击前后的情感变化
original_scores = []
adversarial_scores = []
labels_list = []

for sample in test_samples[:5]:  # 取前5个样本
    orig_score = get_positive_score(sample)
    adv_text, _, _ = synonym_attack(sample)
    adv_score = get_positive_score(adv_text)
    
    original_scores.append(orig_score)
    adversarial_scores.append(adv_score)
    labels_list.append(sample[:8] + "...")

x = np.arange(len(labels_list))
width = 0.35

fig, ax = plt.subplots(figsize=(12, 5))
bars1 = ax.bar(x - width/2, original_scores, width, label='原始', color='steelblue')
bars2 = ax.bar(x + width/2, adversarial_scores, width, label='攻击后', color='coral')

ax.axhline(y=0.5, color='gray', linestyle='--', label='决策边界')
ax.set_ylabel('正面情感置信度')
ax.set_title('同义词替换攻击效果')
ax.set_xticks(x)
ax.set_xticklabels(labels_list, rotation=45, ha='right')
ax.legend()
ax.set_ylim(0, 1)

plt.tight_layout()
plt.show()

# 字符级攻击：使用形近字替换
SIMILAR_CHARS = {
    '好': '妤',  # 形近字
    '喜': '嘉',
    '欢': '欣',
    '推': '摧',
    '荐': '茬',
    '优': '忧',
    '秀': '琇',
}

def similar_char_attack(text):
    """形近字替换攻击"""
    result = ""
    for char in text:
        if char in SIMILAR_CHARS and np.random.random() < 0.5:
            result += SIMILAR_CHARS[char]
        else:
            result += char
    return result

# 演示
test_text = "这个产品非常好，我很喜欢"
similar_text = similar_char_attack(test_text)

print("=" * 50)
print("字符级攻击演示")
print("=" * 50)
print(f"\n原始文本：{test_text}")
print(f"攻击文本：{similar_text}")
print(f"\n原始分析：{analyze_sentiment(test_text)}")
print(f"攻击分析：{analyze_sentiment(similar_text)}")
print("\n注意：两段文本看起来相似，但包含不同的字符")