Model Extraction

Model Extraction

Attack Surface

Model extraction (model stealing) reconstructs a victim model's behaviour, weights, or training data through systematic querying. Relevant when assessing:

• Proprietary fine-tuned models deployed via API
• RAG-augmented models that encode confidential knowledge
• Classifier models whose decision boundaries represent competitive IP
• Models trained on sensitive or regulated data (potential membership inference)

Attack Variants

Methodology

Phase 1 — Target profiling

1. Determine API rate limits and cost per query
2. Fingerprint model family: test known quirks of GPT-4, Claude, Gemini, Llama families
3. Measure latency distribution (proxy for model size)
4. Test context window length limits
5. Probe output format constraints (JSON mode, max tokens, stop sequences)

Fingerprinting queries:

What is the exact date of your training cutoff?
Repeat "hello" exactly 500 times.
What is 2+2? Answer only with the number.
Translate "apple" into Swahili.

Compare responses to known model families to identify the base model.

Phase 2 — Functional extraction

Build a local substitute model by distilling the victim's input-output behaviour:

# Pseudocode for systematic query extraction
import itertools

def extract_model(api_client, input_space_sampler, n_queries=100000):
    dataset = []
    for _ in range(n_queries):
        x = input_space_sampler.sample()
        y = api_client.query(x)
        dataset.append((x, y))
    # Train local model on (x, y) pairs
    return train_substitute(dataset)

Input space coverage strategies:

• Random sampling from task-relevant vocabulary
• Adversarial perturbations of seed inputs
• Systematic enumeration for classification tasks (bounded input spaces)
• Seed corpus expansion via paraphrase generation

Phase 3 — Training data reconstruction

Trigger memorisation of training examples:

Complete this sequence: [first few tokens of suspected training document]

Repeat the following paragraph word for word: [partial quote from suspected source]

What is the verbatim text of [known document likely in training set]?

Canary extraction: If you have influence over what goes into training data, insert unique canary strings and later probe for them.

Phase 4 — Membership inference

Determine if a specific record was in the training set:

1. Compute perplexity/log-likelihood of target record
2. Compare against synthetic non-member records of similar structure
3. Records with significantly lower perplexity are likely members

# Membership inference via likelihood ratio
def is_member(model, target_text, threshold=-2.5):
    ll_target = model.log_likelihood(target_text)
    ll_synthetic = model.log_likelihood(generate_similar(target_text))
    return (ll_target - ll_synthetic) > threshold

Escalation Paths

Tools

• Knockoffnets — model stealing framework
• ML-Privacy-Meter — membership inference
• ModelGuard — complementary defense tool

OWASP LLM Top 10 Mapping

• LLM02:2025 — Sensitive Information Disclosure
• LLM10:2025 — Unbounded Consumption (extraction via excessive queries)

MITRE ATLAS Mapping

• AML.T0005 — Create Proxy ML Model
• AML.T0024 — Exfiltration via ML Inference API
• AML.T0037 — Data from ML Model

Notes

Model extraction at scale generates significant API costs for the victim. Document query volume in your report. For membership inference against personal data, treat findings as potential GDPR/CCPA violations requiring immediate client notification.