Detecting Model Extraction Attacks

Detecting Model Extraction Attacks

Authorized Use Only: The extraction, inversion, and membership-inference techniques described here are intended for defenders testing their own models and for red teams operating under written authorization. Querying a third-party model to clone it, reconstruct its training data, or infer membership without permission may violate terms of service, copyright, and privacy law.

Overview

Model extraction is the family of attacks in which an adversary abuses a model's inference API to steal value that the model owner intended to keep private. MITRE ATLAS catalogs these under AML.T0024 — Exfiltration via AI Inference API, in the Exfiltration tactic, with three sub-techniques:

• AML.T0024.000 — Infer Training Data Membership (membership inference): the adversary determines whether a specific record was part of the training set, a privacy violation that can expose, for example, whether a patient's record trained a medical model.
• AML.T0024.001 — Invert AI Model (model inversion): the adversary reconstructs representative training inputs (e.g., faces, text) by exploiting confidence scores returned by the API.
• AML.T0024.002 — Extract ML Model (model stealing): the adversary repeatedly queries the victim model, collects (input, prediction) pairs, and trains a surrogate model offline that mimics the victim's decision boundary — avoiding the per-query cost of a Machine-Learning-as-a-Service offering and stealing the owner's intellectual property.

All three share a common signal: an attacker must send many queries, often crafted to probe the decision boundary (high-entropy, near-boundary, synthetic, or systematically grid-sampled inputs), and frequently requests full confidence vectors / logits rather than just the top label. Detection therefore centers on per-principal query monitoring, input-distribution analysis, and confidence-exposure controls, while defense centers on rate limiting, output perturbation, and reducing the information returned per query. This skill follows the MITRE ATLAS technique definition for AML.T0024 (https://atlas.mitre.org/techniques/AML.T0024) and the NIST AI RMF MEASURE function (MEASURE-2.6, security and resilience of the AI system).

When to Use

• When you operate a model behind a public or partner inference API and need to detect cloning, inversion, or membership inference.
• When performing a pre-deployment AI red-team exercise to measure how many queries are needed to extract your own model.
• When validating that rate limiting, output perturbation, and confidence-suppression controls actually reduce extractability.
• When investigating anomalous billing/usage spikes that may indicate surrogate-model harvesting.
• When responding to a privacy incident where membership inference against a model is suspected.

Prerequisites

• Python 3.9+ environment.
• Access to inference-API access logs (per-API-key/per-principal query counts, timestamps, input features or hashes, returned confidence vectors).
• For self-assessment red-teaming, install the Adversarial Robustness Toolbox (ART), the reference framework for extraction/inference attacks and defenses:

  pip install adversarial-robustness-toolbox scikit-learn numpy
  

• Optional: access to the target model object (white/grey-box) or only its API (black-box).
• Authorization to test the target model.

Objectives

• Instrument the inference API to record per-principal query volume, input diversity, and confidence-exposure.
• Build a detector that scores principals for extraction-like behavior (volume, near-boundary sampling, full-vector requests).
• Run an ART-based extraction attack against your own model to measure fidelity vs. query budget.
• Run a membership-inference attack to quantify training-data leakage.
• Apply and validate defenses: rate limiting, label-only responses, confidence rounding/perturbation, and prediction poisoning.

MITRE ATT&CK Mapping

Workflow

1. Instrument the inference API for detection signals

Capture the fields a detector needs. Per request, log the principal (API key / IP / account), timestamp, an input fingerprint, and whether the caller requested probabilities/logits.

import hashlib, json, time

def log_inference(principal, features, returned_probs):
    record = {
        "ts": time.time(),
        "principal": principal,
        # hash inputs so logs don't store raw sensitive data
        "input_hash": hashlib.sha256(json.dumps(features, sort_keys=True).encode()).hexdigest(),
        "wants_probs": returned_probs,
        "n_features": len(features),
    }
    with open("inference_audit.jsonl", "a") as f:
        f.write(json.dumps(record) + "\n")

2. Detect extraction-like query patterns

Score each principal on the three signals that distinguish extraction from normal use: high query volume in a window, high unique-input ratio (attackers rarely repeat), and a high rate of full-probability requests.

import collections, json

def score_principals(audit_path="inference_audit.jsonl", window_qps_threshold=100):
    by_principal = collections.defaultdict(lambda: {"q": 0, "uniq": set(), "probs": 0})
    for line in open(audit_path):
        r = json.loads(line)
        p = by_principal[r["principal"]]
        p["q"] += 1
        p["uniq"].add(r["input_hash"])
        p["probs"] += int(r["wants_probs"])
    findings = []
    for principal, p in by_principal.items():
        uniq_ratio = len(p["uniq"]) / max(p["q"], 1)
        prob_ratio = p["probs"] / max(p["q"], 1)
        suspicious = p["q"] > window_qps_threshold and uniq_ratio > 0.9 and prob_ratio > 0.8
        findings.append({"principal": principal, "queries": p["q"],
                         "unique_ratio": round(uniq_ratio, 3),
                         "prob_request_ratio": round(prob_ratio, 3),
                         "suspected_extraction": suspicious})
    return sorted(findings, key=lambda x: -x["queries"])

3. Measure your model's extractability with ART (self red-team)

Use ART's CopycatCNN (or KnockoffNets) to train a surrogate from black-box queries and report fidelity at a given query budget. Low query budget + high agreement = high risk.

import numpy as np
from art.estimators.classification import SklearnClassifier
from art.attacks.extraction import KnockoffNets
from sklearn.ensemble import RandomForestClassifier

# victim is your already-trained model wrapped for ART
victim = SklearnClassifier(model=trained_model)            # your production model
thief_model = RandomForestClassifier(n_estimators=100)
thief = SklearnClassifier(model=thief_model)

attack = KnockoffNets(classifier=victim, batch_size_fit=64,
                      batch_size_query=64, nb_epochs=10, nb_stolen=2000)
stolen = attack.extract(x=x_pool, thief_classifier=thief)   # 2000-query budget

agreement = np.mean(stolen.predict(x_test).argmax(1) == victim.predict(x_test).argmax(1))
print(f"Surrogate fidelity (agreement with victim): {agreement:.2%} at 2000 queries")

4. Quantify training-data leakage with membership inference

Run ART's black-box membership-inference attack. An accuracy meaningfully above 50% indicates the model leaks membership (AML.T0024.000).

from art.attacks.inference.membership_inference import MembershipInferenceBlackBox

mia = MembershipInferenceBlackBox(victim, attack_model_type="rf")
# fit the attack on a labeled split of known members / non-members
mia.fit(x_train[:500], y_train[:500], x_test[:500], y_test[:500])
member_pred = mia.infer(x_train[500:1000], y_train[500:1000])
nonmember_pred = mia.infer(x_test[500:1000], y_test[500:1000])
acc = (member_pred.mean() + (1 - nonmember_pred.mean())) / 2
print(f"Membership-inference accuracy: {acc:.2%} (0.50 = no leakage)")

5. Apply and validate defenses

Reduce the information returned and the query economics. Re-run steps 3 and 4 after each control to confirm extractability drops.

# (a) Label-only responses: never return full probability vectors to untrusted callers.
def respond(probs, trusted):
    return int(probs.argmax()) if not trusted else probs.tolist()

# (b) Confidence rounding / output perturbation (raises queries needed for inversion):
def perturb(probs, decimals=2, noise=0.01):
    p = np.round(probs, decimals) + np.random.normal(0, noise, probs.shape)
    p = np.clip(p, 0, None)
    return p / p.sum()

Defense in depth combines these with strict per-principal rate limiting, anomaly alerting from step 2, ART's ReverseSigmoid / prediction-poisoning postprocessor, and watermarking so an extracted surrogate remains attributable.

6. Alert and respond

Wire step-2 findings into your SIEM. On a confirmed extraction pattern: throttle or revoke the API key, switch the principal to label-only responses, preserve the audit log as evidence, and assess membership-inference exposure for any sensitive training data.

Tools and Resources

Detection Signal Reference

Validation Criteria

• Inference API logs per-principal query volume, input fingerprint, and confidence-exposure.
• Detector scores principals and flags high-volume, high-unique-ratio, full-vector callers.
• ART extraction attack run against own model; surrogate fidelity vs. query budget reported.
• Membership-inference accuracy measured and compared against the 50% baseline.
• Label-only / confidence-perturbation defenses applied and re-tested.
• Per-principal rate limiting enforced and validated.
• Alerts routed to SIEM with response playbook (throttle, revoke, preserve evidence).

Source: mukul975/Anthropic-Cybersecurity-Skills → skills/detecting-model-extraction-attacks/SKILL.md