Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Model Obfuscation
Deliberate transformation process of an AI model to make its internal structure and parameters difficult to interpret by adversaries. Obfuscation protects intellectual property while preserving the model's predictive performance.
Differential Privacy
Mathematical framework ensuring that analysis results do not reveal information about specific individuals in the dataset. This technique adds controlled noise to protect models against inference attacks during their deployment.
Secure Model Transfer Protocol
Set of rules and cryptographic mechanisms ensuring the integrity, authenticity, and confidentiality of models during their transmission. These protocols typically include digital signatures, end-to-end encryption, and integrity verification.
Model Watermarking
Technique of incorporating subtle identification information into a model's parameters to prove intellectual property. Watermarking enables the detection of unauthorized use or rights violation in deployed models.
Adversarial Defense Mechanism
Set of techniques protecting models against adversarial sample attacks during deployment. These mechanisms include robustification, anomaly detection, and defensive randomization.
Encrypted Model Deployment
Deployment architecture where the model remains encrypted during execution in untrusted environments. This approach combines TEE, homomorphic encryption, and secure protocols to preserve complete confidentiality.
Privacy-Preserving Model Publishing
Methodology for sharing pre-trained models while minimizing information leaks about training data. This approach combines compression, private differentiation, and secure approximation techniques.
Secure Inference Protocol
Mechanism ensuring the confidentiality of inputs, outputs, and model parameters during the inference process. These protocols protect against eavesdropping, traffic analysis, and side-channel attacks.
Model Extraction Attack Prevention
Set of countermeasures protecting models from reconstruction by adversaries using inference queries. These techniques include rate limiting, output randomization, and abnormal behavior detection.
Zero-Knowledge Proof
Cryptographic protocol allowing one party to prove knowledge of information without revealing it. In the ML context, ZKP verifies model authenticity without exposing their structure or internal parameters.
Trusted Execution Environment
Secure isolated zone within the processor guaranteeing the confidentiality and integrity of executed code and data. TEEs like Intel SGX or ARM TrustZone enable secure model deployment in shared infrastructures.
Secure Model Compression
Size reduction techniques that preserve model security during the compression process. These methods prevent information leakage while optimizing deployment performance in constrained environments.
Model Serialization Security
Protection measures applied when converting models into storage or transmission formats. Security includes weight encryption, metadata signing, and protection against malicious code injection.
Secure Model Versioning
Version control system integrating cryptographic mechanisms to ensure the integrity and traceability of model evolutions. Each version is signed and hashed to prevent unauthorized modifications.
Model Integrity Verification
Cryptographic validation process confirming that a model has not been altered since its creation or last verification. This verification uses hash functions and digital signatures to ensure trust.
Privacy-Preserving Model Updates
Protocols enabling updates to deployed models without revealing new training data or specific modifications. These approaches combine federated learning and differential masking techniques.