Task-Specific Adaptation with Restricted Model Access
This work addresses inefficiencies and privacy concerns in fine-tuning for users with restricted model access, though it is incremental as it builds on existing fine-tuning paradigms.
The paper tackles the problem of fine-tuning foundational models without accessing their weights, addressing challenges like inefficiency and privacy, by introducing a 'Gray-box' framework with lightweight modules at input and output, achieving competitive performance with full-access methods on benchmarks like text-image alignment.
The emergence of foundational models has greatly improved performance across various downstream tasks, with fine-tuning often yielding even better results. However, existing fine-tuning approaches typically require access to model weights and layers, leading to challenges such as managing multiple model copies or inference pipelines, inefficiencies in edge device optimization, and concerns over proprietary rights, privacy, and exposure to unsafe model variants. In this paper, we address these challenges by exploring "Gray-box" fine-tuning approaches, where the model's architecture and weights remain hidden, allowing only gradient propagation. We introduce a novel yet simple and effective framework that adapts to new tasks using two lightweight learnable modules at the model's input and output. Additionally, we present a less restrictive variant that offers more entry points into the model, balancing performance with model exposure. We evaluate our approaches across several backbones on benchmarks such as text-image alignment, text-video alignment, and sketch-image alignment. Results show that our Gray-box approaches are competitive with full-access fine-tuning methods, despite having limited access to the model.