Glossary

Keyword	Reasoning
Model Fintuning	Model fine-tuning adapts a pre-trained machine learning model for a related task. Instead of training from scratch, this method uses a model that has already learned useful features from another task and fine-tunes it on a new dataset or task.
Model Inference	The model has already been trained and evaluated, and has been deployed into the actual target hardware. Data is input into the model, and the model predicts the results.
Large Language Model/LLM	A type of artificial intelligence model aimed at understanding and generating human language. Trained on a large amount of text data, it can perform a wide range of tasks, including text summarization, translation, sentiment analysis, and so on.
Federated learning	A new paradigm in machine learning aims to train centralized models while keeping a large amount of client-side training data from being uploaded to avoid data leaks.
Deep Gradient Compression/DGC	Reduce the network bandwidth required for exchanging gradient information
Gradient inversion attack	A specialized form of adversarial attack aimed at systems utilizing Federated Learning. In this type of attack, adversaries attempt to 'invert' the aggregated gradient updates sent to the centralized server. By doing so, they seek to reconstruct or infer sensitive information from the original data that contributed to these gradient updates. This poses a significant risk to data exploit and security within Federated Learning ecosystems.
Byzantine attack	Data Poisoning Attack: This attack involves contaminating the local dataset on the client side by adding polluted images. The polluted data causes the client to upload a model with poor accuracy, which in turn affects the global model. Model Poisoning Attack: This type of attack involves altering the model during its training or transmission process, without changing the local data on the client side. Tampering with Controlled Client's Local Data: This attack involves changing the local data on a controlled client device to influence the model's training process.
RAG	Retrieval-Augmented Generation (RAG) is a method that improves the responses of large language models (like GPT-3) by letting them check an external, trustworthy knowledge base before they answer. This means they can give more accurate and up-to-date answers, especially in specialized areas or using a company's own data, without having to be retrained. This makes these models even better at jobs like answering questions, translating languages, and completing sentences, and it's a budget-friendly way to make sure they stay useful and correct in many different situations.