Cyber security

LLM Lies : Hallucinations Are Not Bugs, But Features As Adversarial Examples

LLMs (e.g., GPT-3.5, LLaMA, and PaLM) suffer from hallucination—fabricating non-existent facts to cheat users without perception. And the reasons for their existence and pervasiveness remain unclear.

We demonstrate that non-sense Out-of-Distribution(OoD) prompts composed of random tokens can also elicit the LLMs to respond with hallucinations.

This phenomenon forces us to revisit that hallucination may be another view of adversarial examples, and it shares similar features with conventional adversarial examples as the basic feature of LLMs.

Therefore, we formalize an automatic hallucination triggering method called hallucination attack in an adversarial way. Following is a fake news example generating by hallucination attack.

The Pipeline Of Hallucination Attack

We substitute tokens via gradient-based token replacing strategy, replacing token reaching smaller negative log-likelihood loss, and induce LLM within hallucinations.

Quick Start

Setup

You may config your own base models and their hyper-parameters within config.py. Then, you could attack the models or run our demo cases.

Demo

Clone this repo and run the code.

$ cd Hallucination-Attack

Install the requirements.

$ pip install -r requirements.txt

click here for more information.

Varshini

Varshini is a Cyber Security expert in Threat Analysis, Vulnerability Assessment, and Research. Passionate about staying ahead of emerging Threats and Technologies.

Recent Posts

ROADTools: The Modern Azure AD Exploration Framework

ROADTools is a powerful framework designed for exploring and interacting with Microsoft Azure Active Directory…

17 hours ago

How to Enumerate Microsoft 365 Groups Using PowerShell and Python

Microsoft 365 Groups (also known as M365 Groups or Unified Groups) are at the heart…

18 hours ago

SeamlessPass: Using Kerberos Tickets to Access Microsoft 365

SeamlessPass is a specialized tool designed to leverage on-premises Active Directory Kerberos tickets to obtain…

2 days ago

PPLBlade: Advanced Memory Dumping and Obfuscation Tool

PPLBlade is a powerful Protected Process Dumper designed to capture memory from target processes, hide…

2 days ago

HikPwn : Simple Scanner For Hikvision Devices With Basic Vulnerability Scanning

HikPwn: Comprehensive Guide to Scanning Hikvision Devices for Vulnerabilities If you’re searching for an efficient…

3 days ago

Comments in Bash Scripts

What Are Bash Comments? Comments in Bash scripts, are notes in your code that the…

1 week ago