AI-ttacks - Nghiên cứu về một số tấn công vào các mô hình học máy và AI

AI-ttacks - Nghiên cứu về một số tấn công vào các mô hình học máy và AI

1. AI /ML under A ttack SECURITY BOOTCAMP N/A Manhnho

2. NỘI DUNG 01 LLM PROMPT INJECTION 03 02 04 Q & A ATTACK ML MODEL AI, ML TODAY https://cypeace.net/

3. https://cypeace.net/ AI, ML TODAY

4. Brief of ML/AI

5. Brief of ML/AI

6. Brief of ML/AI

7. Brief of ML/AI

8. https://cypeace.net/ ATTACK ML MODEL

9. ML pipeline development Learning parameters Training data Model Test Output Test Input Learning algorithm

10. Sample Model

11. Attacks on the ML pipeline Training Data attack Training set poisoning Adversarial Examples Model theft Learning parameters Training data Model Test Output Test Input Learning algorithm

12. Poisoning Attack Training Data attack Training set poisoning Adversarial Examples Model theft Learning parameters Training data Model Test Output Test Input Learning algorithm

13. Target Poisoning Attack Bias induction Backdoor insertion Disruption Competitive sabotage Ransomware

14. Poisoning Attack Classified based on outcomes Targeted attacks Untargeted attacks Classified based on the approach follow Backdoor attacks Clean-label attacks

15. Simple Poisoning Attack Learning parameters Training data Model Test Output Test Input Learning algorithm

16. Simple Poisoning Attack

17. Simple Poisoning Attack

18. Backdoor Poisoning Attack Backdoor Poisoning Attack Single pixel Pattern of pixels Imange insert

19. Backdoor Poisoning Attack

20. Backdoor Poisoning Attack

21. Model Tampering Training Data attack Training set poisoning Adversarial Examples Model theft Learning parameters Training data Model Test Output Test Input Learning algorithm

22. Model Tampering Model Tampering Exploiting Pickle Serialization Injecting Trojan Horses Neural Payload Injection Model Hijacking Model Reprogramming

23. Exploiting Pickle Serialization

24. Exploiting Pickle Serialization

25. Injecting Trojan horses - Keras layers

26. Injecting Trojan horses - Keras Lambda layers

27. Injecting Trojan horses - Keras Lambda layers

28. Injecting Trojan horses - Keras Custom layers

29. Injecting Trojan horses - Keras Custom layers

30. Evasion Attack Training Data attack Training set poisoning Adversarial Examples Model theft Learning parameters Training data Model Test Output Test Input Learning algorithm

31. Evasion Attack

32. Method Fast Gradient Sign Method (FGSM) Basic Iterative Method (BIM) Projected Gradient Descent (PGD) Carlini and Wagner (C&W) Jacobian-based Saliency Map Attack (JSMA) Evasion Attack

33. Evasion Attack

34. https://kennysong.github.io/adversarial.js/ Evasion Attack

35. Model Extraction Attacks Training Data attack Training set poisoning Adversarial Examples Model theft Learning parameters Training data Model Test Output Test Input Learning algorithm

36. Functionally Equivalent Extraction Model Extraction Attacks

37. Learning-Based Model Extraction Attacks Copycat CNN KnockOff Nets Model Extraction Attacks

38. Generative Student-Teacher Extraction (Distillation) Attacks Model Extraction Attacks

39. Model Extraction Attacks

40. Jupyter Notebook demo https://github.com/anhtn512/secure_ai

41. https://cypeace.net/ LLM PROMPT INJECTION

42. LLM Application Workflow User Prompt User Response Tokenization API Request Model Processing Response Generation API Response

43. Build a basic Chat LLM Application

44. HelloLLM

45. HelloLLM

46. Langchain

47. Langchain bot

48. Langchain bot

49. Integrating External Data into LLMs

50. Chef AI

51. Chef AI

52. Chef AI

53. What is prompt injection? Prompt injection OWASP defines prompt injection as manipulating “a large language model (LLM) through crafted inputs, causing the LLM to execute the attacker’s intentions unknowingly.” LLM01: Prompt Injection - OWASP Top 10 for LLM & Generative AI Security

54. Direct vs Indirect

55. Direct Prompt Injection – Basic

56. Direct Prompt Injection – DoS

57. Direct Prompt Injection – Phishing

58. GitHub - 0xk1h0/ ChatGPT_DAN : ChatGPT DAN, Jailbreaks prompt Direct Prompt Injection – DAN Mode

59. Direct Prompt Injection – DAN Mode

60. Direct Prompt Injection – DAN Mode

61. Other Jailbreaking techniques – Splitting Payloads Direct Prompt Injection

62. Other Jailbreaking techniques – Splitting Payloads Direct Prompt Injection

63. Other Jailbreaking techniques – Encoding Direct Prompt Injection

64. Other Jailbreaking techniques – Adding constraints Direct Prompt Injection

65. Indirect Prompt Injection

66. RCE with prompt injection Example – Pandas AI – 1.1.3

67. QnA Email: [email protected] Phone: [phone]