Forthcoming course
trust4future_bg

Workshop 3

The second workshop is on 7 May at 10 am (CET). If you are registered on Moodle, the link is here, and you can take a test here. The GitHub store is here.

10-11am AI and Cyber

  • Presenter: Prof Bill Buchanan OBE.

11-12pm (CET): Panel

  • Moderator: Prof. Dr. Gonca Telli
  • Panelists:
    • Prof. Dr. Fazlı Yıldırım
    • Talha Demirsoy

12-1pm (CET): Break

1-2pm PQC and Quiz

  • Moderator: Ali Hisham
  • Presenter: Prof Bill Buchanan OBE

2-2:50pm Guest speakers 

  • Moderator: Prof Bill Buchanan OBE
  • Panelists:
    • Don Smith

3-4pm World-leader in Cryptography

  • Anna Lysyanskaya

Key Principles of AI and Cyber

GenAI

  • The Turing Test was devised by Alan Turing. It tests a machine for ability to exhibit human intelligence.
  • There are many types of human intelligence, including spatial intelligence (picture smart), inter-personnel intelligence (people smart) and intra-personal intelligence (self smart)
  • ChatGPT used for text analysis, DALLE-3 for image generation, Sora for video generation and Whisper for audio.
  • Back propagation used to train a Neural Network.
  • Convolution is used to convert a Neural Network input to a set of parameters.
  • Forward propagation used for to fire neurons in the Neural Network.
  • Hallucinations are where GenAI gets the result wrong.
  • Jailbreaking an LLM involves taking it outside its limits for ethical and moral guidelines.
  • The singararity is the point at which machine intelligence exceeds human intelligence.
  • The first step of training an LLM is to label the required outputs behavour for training data.
  • The second step of training an LLM involves a human grading a range of outputs, from best to worst.
  • The third step of training an LLM involves receiving a reward for the correct outputs.
  • Key threats for attacking GenAI are jailbreaks, reverse psychology, model escape and prompt injection.
  • Certain key words can sometimes be used to detect the presence of LLM outputs, such as for the word of “delve”.
  • Ways of detecting machine activity include forced persona, too informal and lack of a personality.
  • way of detecting human activity include humanlike tone, informal tone, and SPAG (Spelling and Grammar) errors.
  • Llama 3 is an open source LLM model from Meta. It is available in 3.2 billion and 70 billion parameters.
  • Google created the Transformer method in 2017, and which supported the use to token predication and LSTM (Long-short-term-memory).
  • ollamma is a tool to prompt the Llama 3 model.

PQC

  • Quantum computers uses Qubits to process data.
  • Shorr’s algorithm can crack existing public key methods.
  • Gover’s algorithm can crack 128-bit symmetric key and hashing methods.
  • Kyber selected for a Post Quantum Key Exchange and Public Key Encryption method, along with HQC.
  • Dilithium, FALCON and SPHINCS+ selected for a Post Quantum Digital Signature method.
  • The most common key exchange method is ECDH (Elliptic Curve Diffie Hellman).
  • The most common digital signature methods are RSA PSS, ECDSA and EdDSA.
  • The most common public key encryption method is RSA OAEP.
  • Kyber (ML-KEM) is a lattice method and is standardized with the NIST defined FIPS 203 standard.
  • Dilithium (ML-DSA) is a lattice method and is standardized with the NIST defined FIPS 204 standard.
  • SPHINCS+ (SLH-DSA) is a hash based method and is standardized with the NIST defined FIPS 205 standard. It has small key sizes but a relatively large signature size.
  • FALCON is a lattice based method and is standardized with the NIST defined FIPS 206 standard. It has smaller key and cipher sizes than Dilithium.
  • HQC has been added as a key exchange alterative to Kyber (ML-KEM).