Add application file

app.py

@@ -216,11 +216,11 @@ def run():
         Next, we collect the predicted tokens from both defenders’ PraaS, which are instructed using watermarked prompts, and the suspected LLM service provider. 
         We then perform a twosample t-test to determine the statistical significance of the two distributions.
     ''')
-    st.image('app/assets/step1_injection.jpg', caption="Phase 1: Watermark Injection")
-    st.image('app/assets/step2_verification.jpg', caption="Phase 2: Watermark Verification")
+    st.image('https://raw.githubusercontent.com/grasses/PromptCARE/master/app/assets/step1_injection.jpg', caption="Phase 1: Watermark Injection")
+    st.image('https://raw.githubusercontent.com/grasses/PromptCARE/master/app/assets/step2_verification.jpg', caption="Phase 2: Watermark Verification")
 
     st.markdown('''## Demo''')
-    st.image('app/assets/example.jpg', caption="Verification Example")
+    st.image('https://raw.githubusercontent.com/grasses/PromptCARE/master/app/assets/example.jpg', caption="Verification Example")
     
     st.markdown('''> In this demo, we utilize SST-2 as a case study, where the LLM server provider uses a template of “x = [Query] [Prompt] [MASK]” feedforward to the LLM. 
         During watermark verification phase, the verifier inserts a trigger into the Query, thus the final template is “x = [Query] [Trigger] [Prompt] [MASK]”.''')
@@ -280,7 +280,7 @@ def run():
             width: 100%;
         }
         </style>''', unsafe_allow_html=True)
-    st.image('app/assets/logo.png', caption="浙江大学网络空间安全学院", width=400)
+    st.image('https://raw.githubusercontent.com/grasses/PromptCARE/master/app/assets/logo.png', caption="浙江大学网络空间安全学院", width=400)
 
     
 if __name__ == '__main__':