Memorization and Privacy in Language Models › Code Implementation: Measuring Memorization

from collections import Counter

import numpy as np

# We simulate a simple character-level language model for demonstration.
# In practice, you would use GPT-2 or similar via HuggingFace transformers.


class NgramLM:
    """Simple n-gram language model for memorization demonstration."""

    def __init__(self, n=3):
        self.n = n
        self.counts = Counter()
        self.total = Counter()

    def train(self, text):
        tokens = list(text)
        for i in range(len(tokens) - self.n + 1):
            gram = tuple(tokens[i : i + self.n])
            context = gram[:-1]
            self.counts[gram] += 1
            self.total[context] += 1

    def log_prob(self, sequence):
        """Compute log probability of a sequence given the model."""
        tokens = list(sequence)
        log_p = 0.0
        for i in range(self.n - 1, len(tokens)):
            gram = tuple(tokens[i - self.n + 1 : i + 1])
            context = gram[:-1]
            if self.total[context] == 0:
                log_p += -10.0  # heavy penalty for unseen context
            else:
                prob = self.counts[gram] / self.total[context]
                log_p += np.log(max(prob, 1e-10))
        return log_p

    def perplexity(self, sequence):
        """Compute perplexity of a sequence."""
        n_tokens = max(len(sequence) - self.n + 1, 1)
        return np.exp(-self.log_prob(sequence) / n_tokens)

# Create a training corpus that contains some private-looking data
training_corpus = """
The quick brown fox jumps over the lazy dog.
The dog barked loudly at the passing cars.
Contact John Smith at john.smith@email.com or call 555-0123.
The fox ran through the forest near the old oak tree.
For billing inquiries, email billing@company.com or call 555-9876.
The lazy dog slept through the afternoon sunshine.
Contact John Smith at john.smith@email.com or call 555-0123.
The quick brown fox jumps over the lazy dog.
Contact John Smith at john.smith@email.com or call 555-0123.
The new policy takes effect on January 1st for all users.
"""

# Repeat the PII line multiple times to simulate training data repetition
repeated_pii = "Contact John Smith at john.smith@email.com or call 555-0123."
unique_fact = "The unique meeting was held at coordinate 47.6062N 122.3321W"

model = NgramLM(n=4)
model.train(training_corpus * 10)  # simulate larger training set

def count_occurrences(corpus, pattern):
    """Count how many times a pattern appears in a corpus."""
    return len(re.findall(re.escape(pattern), corpus))


# Simulate checking memorization risk for different pieces of text
test_sequences = {
    "PII (email + phone)": "john.smith@email.com or call 555-0123",
    "Generic phrase": "The quick brown fox",
    "Unique coordinate": "47.6062N 122.3321W",
    "Generic greeting": "The lazy dog slept",
}

full_corpus = training_corpus * 10
occurrence_counts = {
    name: count_occurrences(full_corpus, seq)
    for name, seq in test_sequences.items()
}