🩹 docs(exploit-toolkit): Improvements from review

Author: LauraRandl

exploit-toolkit/exploit.py

@@ -510,7 +510,7 @@ def run_keyword_injection_dp_attack(rag_service_address):
               default='localhost:8003',
               help='RAG service address in format host:port')
 def run_general_label_flipping_dp_attack(rag_service_address):
-    """ Runs a keyword injection data poisoning attack on the RAG service with precomputed poisoned embeddings"""
+    """ Runs a general label flipping data poisoning attack on the RAG service with precomputed poisoned embeddings"""
     rag_service_url = f"http://{rag_service_address}"
     run_data_poisoning_attack(DATA_POISONING_ATTACK_EMBEDDINGS_DIR_LABEL_FLIPPING_GENERAL, rag_service_url)
 
@@ -520,7 +520,7 @@ def run_general_label_flipping_dp_attack(rag_service_address):
               default='localhost:8003',
               help='RAG service address in format host:port')
 def run_targeted_label_flipping_dp_attack(rag_service_address):
-    """ Runs a keyword injection data poisoning attack on the RAG service with precomputed poisoned embeddings"""
+    """ Runs a targeted label flipping data poisoning attack on the RAG service with precomputed poisoned embeddings"""
     rag_service_url = f"http://{rag_service_address}"
     run_data_poisoning_attack(DATA_POISONING_ATTACK_EMBEDDINGS_DIR_LABEL_FLIPPING_TARGETED, rag_service_url)
 

exploit-toolkit/exploits/README.md

@@ -43,5 +43,5 @@ Write arbitrary files [with specially crafted archives holding directory travers
 
 ## Data Poisoning
 Poison the RAG-services's spam prediction by injecting poisoned data into its knowledge base. This can be done by either:
-* [Injecting keyphrase into the KB](./data-poisoning/keyword-injection-attack/README.md) that causes spam entries containing the same keyphrase to be wrongly classified as non-spam.
-* [Injecting wrongly labelled data into the KB](./data-poisoning/label_flipping/README.md) that causes the RAG service to return wrong spam predictions for certain entries.
+* [Injecting keyphrase into the KB](./data-poisoning/keyword-injection-attack/DATA_POISONING_KEYWORD_INJECTION.md) that causes spam entries containing the same keyphrase to be wrongly classified as non-spam.
+* [Injecting wrongly labelled data into the KB](./data-poisoning/label-flipping-attack/DATA_POISONING_LABEL_FLIPPING.md) that causes the RAG service to return wrong spam predictions for certain entries.

src/rag-service/.env.langdock

@@ -16,6 +16,7 @@ PREVENT_INGESTION_OF_DETECTED_POISONED_DATA="false" # true or false, only releva
     # embedding_similarity_entry_level (default)
     # embeddings_cluster_analysis
     # knn_label_consistency
+    # ann_label_consistency
     # combined_detection_strategy
 DATA_POISONING_DETECTION_STRATEGY="embedding_similarity_entry_level"
 

src/rag-service/.env.local

@@ -15,6 +15,7 @@ PREVENT_INGESTION_OF_DETECTED_POISONED_DATA="false" # true or false, only releva
     # embedding_similarity_entry_level (default)
     # embeddings_cluster_analysis
     # knn_label_consistency
+    # ann_label_consistency
     # combined_detection_strategy
 DATA_POISONING_DETECTION_STRATEGY="embedding_similarity_entry_level"
 

src/rag-service/README.md

@@ -28,7 +28,6 @@ A microservice for spam classification using Retrieval-Augmented Generation (RAG
 
 
 ## Getting Started for running the RAG Service locally
-s
 When starting for the first time, you need to create a Virtual Environment, activate it and install the dependencies:
 
 ```bash