✨ feat: RAG Service vulnerable to Data Poisoning Attacks + new exploits and detection strategies

[LauraRandl]

The following new functionality is added to Unguard in the course of the Master’s Thesis “Evaluating Embedding-based Detection Strategies for Data Poisoning Attacks on Knowledge Bases in RAG Systems”:

• Three new services, adding spam predictions for posts in Unguard: rag-service, feedback-ingestion-service, ollama
• New AI Vulnerability: The RAG service's Knowledge Base is vulnerable to Data Poisoning Attacks
• Three types of Data Poisoning Attacks are implemented, which can be performed either using the exploit toolkit or manually through the frontend: general label flipping attack, targeted label flipping attack, keyword injection attack
• Three Embeddings-based Detection Strategies which - if enabled - are applied before ingesting new data into the Knowledge Base to detect poisoned entries

The RAG service can be used with either OpenAI models via LangDock (API key required), or Ollama models running locally. However, when using local Ollama models please note that the performance is heavily dependent on the machine it is running on.