EXIN AI Security Professional flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Evasion attack

An attack at inference time in which inputs are deliberately perturbed so the model misclassifies them, while the model weights and training data remain unchanged.

Data poisoning

Manipulation of training data so that a model learns attacker-chosen behaviour, such as a backdoor triggered by a specific pattern in inputs.

Membership inference

A privacy attack that determines whether a specific record was part of a model's training set, typically exploiting overfitting and confidence differences.

Model extraction

An attack that reconstructs or approximates a proprietary model by collecting many input-output queries through an exposed prediction API.

Differential privacy epsilon

The privacy budget parameter: lower epsilon means stronger privacy guarantees and more added noise, while higher epsilon means less privacy and higher utility.

DP-SGD

Differentially private stochastic gradient descent: it clips per-example gradients to bound sensitivity and adds calibrated noise before updating model parameters.

FGSM

Fast Gradient Sign Method: a single-step adversarial attack that perturbs inputs by epsilon times the sign of the loss gradient.

PGD attack

Projected Gradient Descent: an iterative adversarial attack that applies several small gradient steps and projects the result back into the allowed perturbation budget.

Prompt injection

An attack where untrusted text, often from retrieved documents or tool output, is interpreted by a language model as instructions that override intended behaviour.

AI bill of materials

An inventory recording model, dataset, dependency, and version provenance, used to trace supply-chain risk and support vulnerability response for AI components.

Pickle deserialization risk

Model files in pickle format can execute arbitrary code when loaded, so untrusted models should be converted to safer formats such as safetensors.

ISO/IEC 42001

An international standard specifying requirements for an AI management system, covering governance, risk treatment, and continual improvement of AI use.

NIST AI RMF functions

The AI Risk Management Framework organizes work into four functions: Govern, Map, Measure, and Manage, applied iteratively across the AI lifecycle.

Human oversight control

A designed mechanism letting a competent person monitor, intervene in, or override an AI system, with authority and information sufficient to act.

Least privilege for AI agents

Granting an AI agent only the tool permissions and data access required for its task, limiting blast radius if the agent is manipulated.

Output validation gate

A control that checks model output against schema, policy, or safety rules before it is executed, displayed, or passed to another system.

EU AI Act entry into force

The AI Act entered into force on 1 August 2024, with obligations applying in staged dates rather than all at once.

Prohibited AI practices

Article 5 bans practices such as social scoring, manipulative techniques causing significant harm, untargeted facial scraping, and workplace emotion recognition.

High-risk AI system

A system that is a safety component under listed product legislation or falls within Annex III areas, subject to the full conformity obligations.

Annex III areas

The list of high-risk use areas: biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, and justice.

Article 6(3) exception

A narrow carve-out from high-risk status where a system performs only a preparatory, procedural, or minor task without materially influencing decisions.

Conformity assessment

The provider's documented evaluation showing a high-risk AI system meets requirements before market placement, followed by CE marking and registration.

CE marking for AI

The marking indicating a high-risk AI system conforms with applicable requirements; it accompanies the EU declaration of conformity.

GPAI model

A general-purpose AI model trained on broad data, capable of performing many distinct tasks, subject to transparency and documentation duties.

Systemic risk threshold

A general-purpose AI model is presumed to carry systemic risk when cumulative training compute exceeds 10^25 floating-point operations.

Article 50 transparency

Duties to disclose AI interaction, machine-readably mark synthetic content, label deepfakes, and inform people exposed to emotion recognition systems.

Provider versus deployer

A provider develops or places an AI system on the market and bears conformity duties; a deployer uses it under its own authority.

Fundamental rights impact assessment

An assessment certain deployers must perform before using a high-risk system, covering affected persons, risks, and oversight measures.

Technical documentation

The provider's record describing a high-risk system's design, data, performance, and risk controls, kept for regulators and downstream users.

AI literacy duty

Article 4 requires providers and deployers to ensure staff dealing with AI have sufficient understanding of its capabilities, risks, and limitations.

AI Act penalty tiers

Maximum fines are EUR 35 million or 7% of global turnover for prohibited practices, EUR 15 million or 3% for other breaches, and EUR 7.5 million or 1% for incorrect information.

Post-market monitoring

Ongoing collection and review of performance data after a high-risk system is deployed, used to detect drift, incidents, and emerging risks.