About CoEvolution

The contemporary AI landscape demands a holistic framework ensuring security across the supply chain and entire AI lifecycle. Despite existing adversarial attack techniques, a comprehensive end-to-end flow for identifying threats and vulnerabilities with associated risks is lacking. The EU, through initiatives like the AI Act, emphasizes safety and trustworthiness in AI applications but lacks a system managing weaknesses in a networked AI-supply chain. The CoEvolution project integrates its architecture components to create an end-to-end Security, Trust, and Robustness (STR) assessment solution, generating context-aware AI models characterized by their AI Model Bill of Materials (AIMBOM). The goal is a universal hub providing a coherent STR risk assessment and security assurance flow, aligning with MLDevOps and EU AI regulatory frameworks. The paradigm includes novel AI model descriptions, AIMBOM management, security monitoring, and context awareness. CoEvolution introduces a new STR paradigm based on Bills of-Materials, offering a unified approach to describing AI models in supply chains, ensuring STR compliance with EU directives on trust, fairness, data governance, and GDPR guidelines. Open-source trusted datasets and CoEvolution developed AI models enhance the hub’s capabilities, aiming for a robust, adaptable risk analysis and security assessment framework aligned with evolving AI cybersecurity threats.

Architecture

The CoEvolution Hub is a comprehensive framework for enhancing the security, robustness, and trustworthiness of AI models across their lifecycle. It integrates several components to identify and mitigate risks, address vulnerabilities, and enable adaptive responses to dynamic environments.

The AI Risk Assessment Engine evaluates risks related to model design, training, and deployment (e.g., Federated Learning, interconnected AI agents) and recommends mitigation techniques using the CoEvolution Security Trust and Robustness Defense Framework. Complementing this, the Security Testing Engine detects vulnerabilities, such as biases leading to data poisoning, and suggests hardening strategies. Both engines leverage open-source knowledge bases, like the AI Risk Database and MITRE ATLAS, and contribute updates for newly discovered vulnerabilities. Their findings are compiled into Model-Based Bills of Materials (MBOMs), which document the security posture of assessed models.

At its core, the Security Trust and Robustness (STR) Defense Framework offers tools and techniques for each lifecycle phase—design, training, and deployment. These solutions address threats ranging from adversarial attacks to collaborative AI challenges. Defense strategies are realized through “defense flows,” combining complementary techniques across lifecycle stages. In deployment, integrated “gadgets” enable AI models to dynamically detect and adapt to adversarial attacks, enhancing robustness and self-awareness.

The framework also incorporates context-awareness, tailored to the AI architecture (single, collaborative, or interconnected). It uses real-time data and metadata to adapt to operational conditions dynamically, leveraging evolutionary algorithms for complex systems. Attacks detected during runtime trigger alerts to the Security Runtime Monitoring Engine (SRME), which oversees attack surveillance, trust evaluations, and updates or revokes MBOMs if a model is compromised.

The CoEvolution Hub includes a knowledge base of vetted AI models, trusted datasets, and AI MBOMs. These resources are openly accessible to foster ecosystem growth. Orchestrated by a Security Assurance Process, the framework aligns with AI trustworthiness standards, ensuring robust, secure, and context-aware AI deployments.