An question about the code, thank you!

[Kissacat]

Hello! THANK YOU for your work! I do appreciate your contribution

When I read your code in attacker.py, I am confused as to why you concatenate the victim embeddings and attacking embedding in the training phase but only compute the attacking embedding of attacker models in the testing phase. Whether there will be a dimensional mismatch? I'd appreciate it if you could respond to my questions in your busy schedule.

In the training phase (line 236-240 in attacker.py): batch_X = batch_X.to(device) batch_X_unsqueeze = torch.unsqueeze(batch_X, 1) inputs_embeds = torch.cat((batch_X_unsqueeze,input_emb),dim=1) #[batch,max_length+1,emb_dim (1024)]

In the test phase (line 165 in attacker.py): sent_list, gt_list = eval_on_batch(batch_X=embeddings,batch_D=batch_text,model=config['model'],tokenizer=config['tokenizer'],device=device,config=config)

Thank you!!

[teapotliid]

Hi Kissacat, Sorry for the late reply. For mismatched dimensions, we apply a 1-layer neural network (Class: linear_projection) to ensure that the embedding dimensions can be matched.

[Kissacat]

Hi Kissacat, Sorry for the late reply. For mismatched dimensions, we apply a 1-layer neural network (Class: linear_projection) to ensure that the embedding dimensions can be matched.

Thank you!