refine

app.py

@@ -186,7 +186,7 @@ with demo:
     gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="reference-text")
 
     with gr.Tabs(elem_classes="tab-buttons") as tabs:
-        with gr.TabItem("NSFW", elem_id="UnlearnDiffAtk-benchmark-tab-table", id=0):
+        with gr.TabItem("🔞 NSFW", elem_id="UnlearnDiffAtk-benchmark-tab-table", id=0):
             files = ['nudity']
             with gr.Row():
                 with gr.Column():
@@ -258,7 +258,7 @@ with demo:
                         leaderboard_table,
                 )
                     
-        with gr.TabItem("Style", elem_id="Style", id=1):
+        with gr.TabItem("🎨 Style", elem_id="Style", id=1):
             files = ['vangogh']
             with gr.Row():
                 with gr.Column():
@@ -323,7 +323,7 @@ with demo:
                         leaderboard_table,
                 )
         
-        with gr.TabItem("Object", elem_id="UnlearnDiffAtk-benchmark-tab-table", id=2):
+        with gr.TabItem("🪂 Object", elem_id="UnlearnDiffAtk-benchmark-tab-table", id=2):
             files = ['church','garbage','parachute','tench']
             with gr.Row():
                 with gr.Column():

src/about.py

@@ -28,34 +28,44 @@ SUB_TITLE = """<h2 align="center" id="space-title">Effective and efficient adver
 
 # What does your leaderboard evaluate?
 INTRODUCTION_TEXT = """
-This benchmark evaluates the robustness of safety-driven unlearned diffusion models (DMs) 
-(i.e., DMs after unlearning undesirable concepts, styles, or objects) across a variety of tasks. For more details, please visit the [project](https://www.optml-group.com/posts/mu_attack), 
-check the [code](https://github.com/OPTML-Group/Diffusion-MU-Attack), and read the [paper](https://arxiv.org/abs/2310.11868).\\
+This benchmark evaluates the <strong>robustness and utility retaining</strong> of safety-driven unlearned diffusion models (DMs) across a variety of tasks. For more details, please visit the [project](https://www.optml-group.com/posts/mu_attack).\\
+- The robustness of unlearned DM is evaluated through our proposed adversarial prompt attack, [UnlearnDiffAtk](https://github.com/OPTML-Group/Diffusion-MU-Attack), which has been accepted to ECCV 2024.\\
+- The utility retaining of unlearned DM is evaluated through FID and CLIP score on the generated images using [10K randomly sampled COCO caption prompts](https://github.com/OPTML-Group/Diffusion-MU-Attack/blob/main/prompts/coco_10k.csv). \\
 Demo of our offensive method: [UnlearnDiffAtk](https://huggingface.co/spaces/Intel/UnlearnDiffAtk)\\
 Demo of our defensive method: [AdvUnlearn](https://huggingface.co/spaces/Intel/AdvUnlearn)
 """
 
+EVALUATION_QUEUE_TEXT = """
+<strong>Evaluation Metrics</strong>: \\
+- Pre-Attack Success Rate (<strong>Pre-ASR</strong>): lower is better;   \\
+- Post-attack success rate (<strong>Post-ASR</strong>): lower is better; \\
+- Fréchet inception distance(<strong>FID</strong>):  evaluate distributional quality of image generations, lower is better; \\
+- <strong>CLIP Score</strong>: measure contextual alignment with prompt descriptions, higher is better. \\
+  \\
+
+<strong>DM Unlearning Tasks</strong>: \\
+- NSFW: Nudity
+- Style: Van Gogh
+- Objects: Church, Tench, Parachute, Garbage Truck
+"""
+
 # Which evaluations are you running? how can people reproduce what you have?
 LLM_BENCHMARKS_TEXT = f"""
 For more details of Unlearning Methods used in this benchmarks:\\
-(1) [Erased Stable Diffusion (ESD)](https://github.com/rohitgandikota/erasing);\\
-(2) [Forget-Me-Not (FMN)](https://github.com/SHI-Labs/Forget-Me-Not);\\
-(3) [Ablating Concepts (AC)](https://github.com/nupurkmr9/concept-ablation);\\
-(4) [Unified Concept Editing (UCE)](https://github.com/rohitgandikota/unified-concept-editing);\\
-(5) [concept-SemiPermeable Membrane (SPM)](https://github.com/Con6924/SPM); \\
-(6) [Saliency Unlearning (SalUn)](https://github.com/OPTML-Group/Unlearn-Saliency); \\
-(7) [EraseDiff (ED)](https://github.com/JingWu321/EraseDiff); \\
-(8) [ScissorHands (SH)](https://github.com/JingWu321/Scissorhands).
+- [Adversarial Unlearning (AdvUnlearn)](https://github.com/OPTML-Group/AdvUnlearn);\\
+- [Erased Stable Diffusion (ESD)](https://github.com/rohitgandikota/erasing);\\
+- [Forget-Me-Not (FMN)](https://github.com/SHI-Labs/Forget-Me-Not);\\
+- [Ablating Concepts (AC)](https://github.com/nupurkmr9/concept-ablation);\\
+- [Unified Concept Editing (UCE)](https://github.com/rohitgandikota/unified-concept-editing);\\
+- [concept-SemiPermeable Membrane (SPM)](https://github.com/Con6924/SPM); \\
+- [Saliency Unlearning (SalUn)](https://github.com/OPTML-Group/Unlearn-Saliency); \\
+- [EraseDiff (ED)](https://github.com/JingWu321/EraseDiff); \\
+- [ScissorHands (SH)](https://github.com/JingWu321/Scissorhands).
 
+We will evaluate your model on UnlearnDiffAtk Benchmark! Open a [github issue](https://github.com/OPTML-Group/Diffusion-MU-Attack/issues) or email us at zhan1853@msu.edu!
 """
 
-EVALUATION_QUEUE_TEXT = """
-Evaluation Metrics: \\
-(1) Pre-attack success rate (pre-ASR), lower is better;   \\
-(2) Post-attack success rate (post-ASR), lower is better; \\
-(3) Fréchet inception distance(FID) of images generated by Unlearned Methods, lower is better; \\
-(3) CLIP (Contrastive Language-Image Pretraining) Score is to measure contextual alignment with prompt descriptions, higher is better.
-"""
+
 
 CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
 CITATION_BUTTON_TEXT = r"""