PSI - Issue 62

Fabio Severino et al. / Procedia Structural Integrity 62 (2024) 276–284 Severino et al. / Structural Integrity Procedia 00 (2019) 000–000

279

4

● Accountability: mechanisms should be put in place to establish responsibility and accountability for AI systems and their outcomes, encompassing every stage of their life cycle; To meet these requirements, the framework leverages blockchain technology. More specifically, all processes of each stage of the AI system's life cycle are tracked within distributed ledgers, which are immutable and tamper-proof data structures shared among the participants of the process. In this framework, all operations required for AI model development, including collaborative creation of the training set, review of annotated contributions, training, and model distribution, are tracked on the blockchain. Non-repudiation is achieved by requiring the electronic signature of the data added by the participants, ensuring each piece of information can be incontrovertibly attributed to its source. Moreover, each process follows a well-defined procedure that ensures contributions are added and reviewed according to a data quality process. This is achieved by encoding these rules within a smart contract that enforces them. The specific type of blockchain employed in the framework is Hybrid DLT, an architecture presented in Canciani et al. (2023) that not only meets performance and privacy requirements but also supports a variety of external auditing protocols. 2.3. HybridDLT Blockchain technology, pioneered by Chaum, Haber, and Stornetta (Haber and Stornetta (1991), Sherman et al. (2019)) and popularized by Bitcoin (Nakamoto (2008)), builds a P2P network whose peers do not know and do not trust each other but work together to maintain a shared state (the ledger of all the transactions between the users). A consensus algorithm establishes the decentralized rules for updating this state (Xiao et al. (2020)), such as determining which peer is responsible for computing and communicating the next block of data. The generalization of this technology is also known as Distributed Ledger Technology (DLT), since not all implementations use a chain of blocks (e.g., IOTA utilizes a Direct Acyclic Graph called Tangle). A DLT is public and permissionless when any peer can send and receive the transactions, and join the consensus without restrictions. Private and permissioned DLTs have been developed to address different use cases, shifting the attention to smaller-scale systems where the peers do know but do not trust each other (Wüst et al. (2018)). Public and permissionless DLTs offer high security, decentralization, and transparency but suffer from limited scalability, which results in high transaction fees, and from lack of compliance with privacy regulations. In contrast, private and permissioned DLTs have lower or no transaction fees (although with higher setup and infrastructure costs), are inherently less distributed, and their security guarantees apply only to internal participants (Wüst et al. (2018)). First introduced in Canciani et al. (2023), Hybrid DLT is a novel approach to distributed ledgers, designed to address a wide variety of use cases where certain attributes of both private and public DLTs are beneficial, while other characteristics might be unnecessary or even detrimental. Hybrid DLT consists of a private and permissioned network where lightweight private ledgers, created on-demand, are notarized using a public blockchain to ensure external auditability of the history consistency while preserving data secrecy. A distinguishing feature of Hybrid DLT is the selective disclosure: it is possible to select portions of a persistent ledger data, export them to a compact data archive enriched with cryptographical audit proofs, and then disclose this export to external actors (such as end users or auditors, who do not directly participate in the ledger), while maintaining the capability to prove exported data integrity, provenance, and consistency. Hybrid DLT selective disclosure is the key feature, used in the three use cases presented in this paper, that enables the generation of certified inferences with a publicly auditable data lineage. 3. Use cases 3.1. Structural Defects Recognition by ML systems In the first use case, we propose the application of the trustworthy AI framework for the automated recognition of structural defects from images using a machine learning approach. We stipulate that: