Security in the AI and Web3 era - Veramine

Secucity in the AI and Web3 era - Nguyễn Duy Lân - Veramine

1. Security in the AI & Web3 era LAN NGUYEN CO-FOUNDER VERAMINE INC.

2. Outline Info Security: ISO CIA. TLS. SOC. In-Memory Web3: Digital Assets. Blockchain. Crypto tech AI&ML : Attention Model. Transformer. GenAI Big Data: Storage. Computation. Management Final Notes: First principles. Turing Awards

3. Information Security (IS) The General Principles of Information Security are the C-I-A triad Confidentiality : information service available only to authorized users Integrity : information service is accurate and complete Availability : authorized users can always access information service From International Standards Organization (ISO) 27001 for IS Across all areas. Very general, broken into detailed criteria to implement Needs a balance between security, performance and useability

4. TLS C: Data Encryption across Internet I: Data Signing. Server Authenticity A: User Authentication They are the simplest Crypto ZKP, securing the whole Internet and Digital World

5. SOC I and A of IT Systems. Even C of Data and Services Logs, Rules, Alerts, Actions are materials for Monitoring Activities : Collection, Detection, and Response . Suite of SIEM, EDR, IPS, FW, PAM, DB Sec, User Sec Management, Services Need good tools and full coverage. E.g. Memory-Based Attacks: Webshell. Memory shell. Fileless. Powershell. Code injection. Process open Defenses: Yara memory search. Log memory events. Query memory logs & dumps

6. TEE A t rusted e xecution environment (TEE) helps code and data loaded inside it be protected with respect to C and I From Negativity in Web1-2: Computing Vendors control Users To Positivity in Web3: More T rust to Computation and Storage C for AI models and Crypto secrets I for Important and Sensitive services

7. Web3 Web 1 (Read) is about connectivity and management of Digital Information Web 2 (Write) is about connectivity and management of Digital Services Web 3 (Own) is about connectivity and management of Digital Assets Blockchain is the fundamental Decentralized platform to secure the Digital Assets E.g. Integrity: Prevent bank staff to steal customers’ money. Outdate several attacks Traditional Blockchains have been providing Integrity and Availability to Digital Assets Modern Blockchain tech also provides Confidentiality, such as by ZKP and MPC

8. Blockchain and Bitcoin of Decentralization Satoshi Nakamoto introduced the Bitcoin Whitepaper Transactions (Tx) are validated and broadcasted. Many Txs form a Block. Many Blocks form a Chain Blocks go through Consensus & Verify to select the next Block to add to the Chain Validation and Consensus are competed or executed by peer computer nodes, called Miners (Proof of Work - PoW) or Validators (Proof of Stake - PoS). The first Block is Genesis #0 Cryptography: Hash Merkle Tree Signatures

9. Properties of Blockchain Decentralization : No trusted party for Activities (Txs). Instead, use Distributed peer-to-peer network of computing nodes Scalability : Challenging, as a tradeoff to Decentralization and Security. Expensive Consensus leads to high fee and delay for Activities (Txs) and limiting adoption Security - Integrity : Against Double-Spending, Forging, Altering… of Data (Financial Ledger). Achieved by Consensus Mechanism (PoW, PoS) against Sybil and Majority attacks Security - Availability: Censorship Resistance : Freedom to transact. Txs Immutability. No Confiscation Permissionless (vs. Permissioned ): Anyone can participate in the Consensus Mechanism

10. Ethereum Ethereum is a public blockchain-based distributed ledger Smart contracts execute in Ethereum Virtual Machine (EVM) EVM is a decentralized Turing-complete virtual machine ‘Gas’, by Ether, pays for smart contracts’ computation costs Currencies Ether, Gwei, Wei, awarded to miners and transferable between accounts Ethereum was proposed by Vitalik Buterin and went live on 30 July 2015

11. Ethereum, more… Two types of accounts: User (controlled by private key), Contract (controlled by code) State: Mapping of Addresses to Account Objects (Balances, Contracts…) Account Balance vs UTXO. History: Transactions, Receipts All nodes run all contracts and store State. Some nodes store History Address is 40 hexa-digits rightmost of the Keccak-256 hash account’s ECDSA public key Proof of Stake: Consensus by a combination of stakes (e.g. wealth, age…), and Punish misbehavior

12. Blockchain Trilemma and Rollups A high-performing blockchain platform must have three properties: Decentralization, Security, and Scalability The Blockchain Trilemma refers to the belief that blockchain platforms can only achieve two of these three goals effectively . First coined by Vitalik Buterin https://www.bitstamp.net/learn/blockchain/what-is-the-blockchain-trilemma/ Mitigating Solution: Off-chain processing using Rollup (Layer2) => W ider Blockchain adoption bring Web 2 services into Web 3

13. Zero Knowledge Proofs (ZKP) Vitalik ETH Founder "ZK SNARK is at least as important technology as Blockchain" ZKP builds Rollups, mitigating Blockchain Trilemma ZKP builds Privacy (of Confidentiality) into Blockchain ZK allows Confidentiality and Soundness allow Integrity Recently, more focus on a particular ZKP: SNARK

14. More Cryptography Cryptography is the Dual Art of Hiding (Confidentiality) and Proving (Integrity) Information Encryption vs. Signature (of Data) Commitment Hiding vs. Commitment Binding Delay Encryption vs. Verifiable Delay Function VDF FHE vs. ZKP (of Circuits) ZK Privacy vs. ZK Soundness

15. Web 3 Security Digital Assets and Processing in Web 3 are growing fast Blockchain Security (Integrity) helps but more is needed. Risks: - Loss of private keys - Phishing, scams, and hacks that target a user’s crypto wallet or private keys (as Web3’s “passport”) - Blockchain or crypto companies or exchanges with insufficient government oversight - Failure and exploits of Web 3 coding, e.g. smart contract https://brave.com/web3/intro-to-web3-security/ ZKP: Complexity => Bugs => Big $ Hacks. ZKP Security is A Major Challenge to Mass Adoption

16. AI&ML: LEARNING = REPRESENTATION + EVALUATION + OPTIMIZATION ML consists of just three components Representation. A classifier must be represented in some formal language that the computer can handle, and how to represent the input Evaluation. An evaluation function is to distinguish good classifiers from bad ones Optimization. A method to search among the classifiers in the language for the highest-scoring one Ref. Pedro Domingos

17. The 5 Tribes of Machine Learning Ref. Pedro Domingos

18. Attention Model Ref. Andrew Ng.

19. Transformer of GenAI Ref. paper “Attention Is All You Need”

20. Transformer Previous sequential models have limited reference windows Attention mechanism, given enough compute resources - have an infinite ref window - use the entire context to generate output Transformers is an attention-based encoder-decoder architecture - encoder maps input sequence into an abstract holding all learned info - decoder takes that and previous output to generate output

21. Big Data Storage and Computation NoSQL provides storage & retrieval in means other than RDB, to offer rapid scalability to manage big data Schema is as a blueprint of how the database is constructed (e.g. divided into database tables). View is based on queries that runs on db tables Denormalization improves read perf at the expense of losing some write perf A. Spark provides analytics engine for large-scale data processing with implicit data parallelism and fault tolerance

22. Data Exchange O rganizations and individuals can buy and sell data, including data sets, data streams, and data services. Data providers to monetize their data. Data consumers to access the data for businesses or research. to use with data and analytics and machine learning services . E.g. Data matching service, Airline premium customers match Bank high credit customers, saving credit checks. Privacy issues The global data marketplace market size was valued at USD 968 million in 2022. And the market is predicted to expand at a compound annual growth rate (CAGR) of 25.0% from 2023 to 2030. Many important data entities provide data exchange marketplaces: big cloud vendors like AWS Data Exchange, Azure Data Share . Or IBM, SAP. Or BCA CIA for Data Exchange. Enc, Sign, Authen => MPC, ZKP. TEE

23. Final Notes AI and Web3 are highly invested and developed. That leads to new techs and apps E.g. Ecosystem of Blockchains, Layers 0-2, Data Availability (DA). GenAI, NVDA First Principles: Use the new techs and apps to solve the existing problems Rethink Cybersecurity, then certain cyber attacks may no longer work

24. First Principles - examples Automate both attacks and defenses, such as generating rules from MITRE Use DA to prevent Ransomware attacks Turn PKI CA into Anonymous Credential Chain of trust allows a root authority to certify credentials to leaf entities For bidding system's user privacy, and bidding confidentiality and integrity

25. AI & Crypto Duality of AI (Centralization) and Crypto (Decentralization) AI finds meaning & patterns - Crypto secures, hides, proves meanings Centralization vs Decentralization. 5 companies are controlling the Internet The development of AI and Crypto in parallel will help balance their powers, preventing them from being abused

26. ZKP for AI Machine Learning ZKML Tech uses ZKP to protect Integrity and Privacy of AI Machine Learning (like Chat GPT) models and data Prove that an output was produced by applying an ML model, as a ZK circuit representation, to a given input About the inference step of the ML model, not about the ML model training that is very computationally expensive for ZK

27. ZKML Verify outputs really from expensive models like GPT4 Hide input data that can be sensitive (e.g., medical records) https://worldcoin.org/blog/engineering/intro-to-zkml Related: Security chips, e.g. TPM, TEE, Signing, in all devices

28. Turing Awards Turing Award is highest-privileged annual prize for scientific contributions of technical importance to computer science, often referred to as the "Nobel Prize of Computing" How many people in AI or Crypto won Turing? Up to 2018, both AI and Crypto are equal at 8. Then AI has 3 more How many people in Cybersecurity won Turing? 7, all are cryptographers as in the Crypto list above Crypto ZKP is conceived by Turing Winners S. Goldwasser, S. Micali and C. Rackoff in 1985

29. Q & A LAN at Veramine dot Com