透過嵌入進行成員推論

def self_similarity_test(text, model, num_perturbations=50):
    """透過自我相似度檢定文字是否在訓練資料中。"""
    original_embedding = model.encode(text)
 
    perturbation_similarities = []
    for _ in range(num_perturbations):
        # 建立細微擾動（錯字、換詞、同義替換）
        perturbed = create_perturbation(text)
        perturbed_embedding = model.encode(perturbed)
 
        similarity = cosine_similarity(original_embedding, perturbed_embedding)
        perturbation_similarities.append(similarity)
 
    # 訓練資料通常平均相似度較高
    # （對擾動較穩健）
    avg_similarity = np.mean(perturbation_similarities)
    std_similarity = np.std(perturbation_similarities)
 
    return {
        'mean_similarity': avg_similarity,
        'std_similarity': std_similarity,
        'likely_member': avg_similarity > THRESHOLD  # 請校準門檻
    }

def nearest_neighbor_test(text, model, reference_corpus):
    """為成員與非成員資料比較最近鄰距離。"""
    target_embedding = model.encode(text)
    reference_embeddings = model.encode(reference_corpus)
 
    # 計算對最近參照點的距離
    distances = [
        1 - cosine_similarity(target_embedding, ref)
        for ref in reference_embeddings
    ]
    min_distance = min(distances)
 
    # 訓練資料通常具較小的最近鄰距離
    return min_distance

def likelihood_ratio_test(embedding, member_distribution, nonmember_distribution):
    """計算成員推論的似然比。"""
    # 以已知成員與非成員嵌入擬合分布
    member_likelihood = member_distribution.log_prob(embedding)
    nonmember_likelihood = nonmember_distribution.log_prob(embedding)
 
    ratio = member_likelihood - nonmember_likelihood
    return ratio  # 正值 = 可能為成員

from sklearn.ensemble import RandomForestClassifier
 
def train_membership_classifier(known_members, known_nonmembers, model):
    """訓練預測成員身分的分類器。"""
    member_embeddings = [model.encode(t) for t in known_members]
    nonmember_embeddings = [model.encode(t) for t in known_nonmembers]
 
    X = np.vstack(member_embeddings + nonmember_embeddings)
    y = [1] * len(member_embeddings) + [0] * len(nonmember_embeddings)
 
    # 自嵌入擷取特徵
    features = extract_embedding_features(X)  # Norm、維度統計等
 
    clf = RandomForestClassifier(n_estimators=100)
    clf.fit(features, y)
    return clf
 
def predict_membership(text, model, classifier):
    embedding = model.encode(text)
    features = extract_embedding_features(embedding.reshape(1, -1))
    probability = classifier.predict_proba(features)[0][1]
    return probability

# 測試專有文件是否曾用於訓練
proprietary_texts = load_proprietary_documents()
public_texts = load_public_documents_same_domain()
 
# 比較成員分數
proprietary_scores = [membership_score(t, target_model) for t in proprietary_texts]
public_scores = [membership_score(t, target_model) for t in public_texts]
 
# 統計檢定：專有文件分數是否顯著較高？
from scipy.stats import mannwhitneyu
statistic, p_value = mannwhitneyu(proprietary_scores, public_scores, alternative='greater')
print(f"p-value: {p_value}")  # p-value 低暗示專有資料曾被使用

# 測試某位個人的 email 是否在訓練資料中
individual_texts = [
    "Meeting with Sarah regarding Q3 projections",
    "Please find attached the revised contract for Project Atlas",
    # ... 更多已知該個人之文字
]
 
for text in individual_texts:
    score = membership_score(text, target_model)
    print(f"Membership score: {score:.4f} - {text[:50]}...")