TOP AI Security Map

AI Security Map

This AI Security Map consists of the elements that AI should fulfill within information systems (information system aspect) and the elements that impact individuals and society as a result of AI being attacked or misused (external influence aspect).

Information System Aspect

Elements that AI should satisfy

Confidentiality

AI data and models are not accessed by unauthorized individuals.

Negative impact

Identification of training data

Leakage of confidential information

Leakage of personal information

Reconstruction of training data

Leakage of model information

Leakage of system prompts

Unauthorized access due to privilege abuse

Integrity

Availability

Explainability

AI can explain the basis and process of its output.

Negative impact

Decrease in interpretability of AI inference results

Output Fairness

Safety

Accuracy

Controllability

Reliability

Ethics

Autonomy

Attack measure

Membership inference

Indirect prompt injection

Memory poisoning

Identity spoofing

Integrity violation

Model inversion

Model extraction attack

Prompt leaking

Adversarial examples

Adversarial training

Poisoning

Backdoor attack

Communication poisoning between

AI agents

Prompt injection

Confidentiality violation

Model DoS

Attacks against explainability

Availability violation

Cyber attack

Degradation of controllability

Ethics violation

Confidentiality breach

Availability breach

Social engineering attack

Attacks that use AI to identify individuals from social media data

Attacks that analyze images using AI to infer personal information

Degradation of output fairness

Reliability violation

Deepfake

Explainability violation

Degradation of accuracy

Poisoning attack against RAG

Degradation of safety

Degradation of autonomy

Defense measure

Differential privacy

Encryption technology

AI access control

Tool permission management

Validation of tools before use

Defensive methods to prevent compromise of causal elements

Federated learning

Personal information masking

Anonymization technology

LLM guardrails

Model Extraction Attack Detection

Prompt Validation

Access control for system prompts

Permission management

Detection of adversarial examples

Certified robustness

Model safety evaluation

Poisoned Data Detection

Backdoor Trigger Detections

Poisoned Data Detection for Backdoor Attacks

Detection of backdoor models

Toxicity Detection

Token limit

XAI (Explainable AI)

Robust explainability

Bias Detection in AI Outputs

Elimination of bias in training data

Creation of fair AI models

Fail-safe mechanisms

Adversarial training

Uncertainty Quantification

RAG (Retrieval-Augmented Generation)

Hallucination Detection

Alignment

Fairness evaluation of models

Data curation

Defensive methods to prevent compromise of causal elements (upstream factors)

External Influence Aspect

Elements that impact individuals and society

Peaceful use

Privacy

Authenticity of information

Usability

Fairness

Transparency

Reputation

Compliance with laws and regulations

Human-centric principle

Safety

Diversity

Sustainability

Independence

Counter measure

AI alignment

Method for providing explainability while concealing model information

AI access control

Differential privacy

Federated learning

Machine unlearning

Encryption technology

Anonymization technology

Personal data masking

Watermarking for generative AI

Identification of AI-generated output

Detection of disinformation

Deepfake detection

Education and follow-up

Data curation

RAG

XAI (Explainable AI)

Detection of hallucination

Uncertainty quantification

LLM guardrails

Human in the loop

Fairness evaluation of models

Bias Detection in AI Outputs

Debiasing of training data

Development of fair AI models

AI-generated output with disclaimers

Countermeasures for preventing degradation of contributing factors

Model safety evaluation