AIbase
EN
Home
FARE4 ViT B 16 Laion2b S34b B88k
F

FARE4 ViT B 16 Laion2b S34b B88k

Developed by chs20
A robust perceptual metric model based on CLIP, enhanced through adversarial fine-tuning for improved performance in perceptual similarity tasks.
Image-to-Text
Safetensors
Open Source License:MIT#Adversarially Robust CLIP#Zero-shot Image Classification#Perceptual Similarity Metric
Downloads 23
Release Time : 8/14/2024

Model Overview

This model is a vision-language model based on the CLIP architecture, adversarially fine-tuned (using the FARE method) to significantly improve robustness against adversarial attacks while maintaining performance on clean data. Primarily used for zero-shot image classification and perceptual similarity tasks.

Model Features

Adversarial Robustness
Adversarially fine-tuned on ImageNet with L-infinity norm and radius 4/255 using the FARE method, significantly enhancing the model's resistance to adversarial attacks.
High-performance Perceptual Metric
Excels in the NIGHTS perceptual similarity task, achieving 90.6% accuracy on clean data and maintaining high performance under adversarial attacks.
CLIP-based Architecture
Built on the mature CLIP-ViT-B-16 architecture, inheriting CLIP's strong vision-language alignment capabilities.

Model Capabilities

Zero-shot Image Classification
Perceptual Similarity Metric
Adversarially Robust Image Analysis

Use Cases

Computer Vision
Image Similarity Assessment
Used to evaluate the perceptual similarity between two images from a human perspective.
Achieves 90.6% accuracy on the NIGHTS dataset.
Robust Image Classification
Maintains good classification performance even in the presence of adversarial interference.
Maintains 71.5% accuracy under L-infinity attack (eps=4/255).
Safety-critical Applications
Adversarial Attack Detection
Identifies image content that may have been tampered with by adversarial attacks.
Featured Recommended AI Models
AIbase
Empowering the Future, Your AI Solution Knowledge Base
EnglishįŽ€äŊ“中文įšéĢ”中文ãĢãģんご
Š 2025AIbase