import os import zipfile import gradio as gr import nltk import pandas as pd import requests from pyabsa import TADCheckpointManager from textattack.attack_recipes import ( BAEGarg2019, PWWSRen2019, TextFoolerJin2019, PSOZang2020, IGAWang2019, GeneticAlgorithmAlzantot2018, DeepWordBugGao2018, CLARE2020, ) from textattack.attack_results import SuccessfulAttackResult from utils import SentAttacker, get_agnews_example, get_sst2_example, get_amazon_example, get_imdb_example, diff_texts # from utils import get_yahoo_example sent_attackers = {} tad_classifiers = {} attack_recipes = { "bae": BAEGarg2019, "pwws": PWWSRen2019, "textfooler": TextFoolerJin2019, "pso": PSOZang2020, "iga": IGAWang2019, "ga": GeneticAlgorithmAlzantot2018, "deepwordbug": DeepWordBugGao2018, "clare": CLARE2020, } def init(): nltk.download("omw-1.4") if not os.path.exists("TAD-SST2"): z = zipfile.ZipFile("checkpoints.zip", "r") z.extractall(os.getcwd()) for attacker in ["pwws", "bae", "textfooler", "deepwordbug"]: for dataset in [ "agnews10k", "sst2", "MR", 'imdb' ]: if "tad-{}".format(dataset) not in tad_classifiers: tad_classifiers[ "tad-{}".format(dataset) ] = TADCheckpointManager.get_tad_text_classifier( "tad-{}".format(dataset).upper() ) sent_attackers["tad-{}{}".format(dataset, attacker)] = SentAttacker( tad_classifiers["tad-{}".format(dataset)], attack_recipes[attacker] ) tad_classifiers["tad-{}".format(dataset)].sent_attacker = sent_attackers[ "tad-{}pwws".format(dataset) ] cache = set() def generate_adversarial_example(dataset, attacker, text=None, label=None): """if not text or text in cache: if "agnews" in dataset.lower(): text, label = get_agnews_example() elif "sst2" in dataset.lower(): text, label = get_sst2_example() elif "MR" in dataset.lower(): text, label = get_amazon_example() # elif "yahoo" in dataset.lower(): # text, label = get_yahoo_example() elif "imdb" in dataset.lower(): text, label = get_imdb_example()""" cache.add(text) result = None attack_result = sent_attackers[ "tad-{}{}".format(dataset.lower(), attacker.lower()) ].attacker.simple_attack(text, int(label)) if isinstance(attack_result, SuccessfulAttackResult): if ( attack_result.perturbed_result.output != attack_result.original_result.ground_truth_output ) and ( attack_result.original_result.output == attack_result.original_result.ground_truth_output ): # with defense result = tad_classifiers["tad-{}".format(dataset.lower())].infer( attack_result.perturbed_result.attacked_text.text + "$LABEL${},{},{}".format( attack_result.original_result.ground_truth_output, 1, attack_result.perturbed_result.output, ), print_result=True, defense=attacker, ) if result: classification_df = {} classification_df["is_repaired"] = result["is_fixed"] classification_df["pred_label"] = result["label"] classification_df["confidence"] = round(result["confidence"], 3) classification_df["is_correct"] = str(result["pred_label"]) == str(label) advdetection_df = {} if result["is_adv_label"] != "0": advdetection_df["is_adversarial"] = { "0": False, "1": True, 0: False, 1: True, }[result["is_adv_label"]] advdetection_df["perturbed_label"] = result["perturbed_label"] advdetection_df["confidence"] = round(result["is_adv_confidence"], 3) advdetection_df['ref_is_attack'] = result['ref_is_adv_label'] advdetection_df['is_correct'] = result['ref_is_adv_check'] else: return generate_adversarial_example(dataset, attacker) return ( text, label, result["restored_text"], result["label"], attack_result.perturbed_result.attacked_text.text, diff_texts(text, text), diff_texts(text, attack_result.perturbed_result.attacked_text.text), diff_texts(text, result["restored_text"]), attack_result.perturbed_result.output, pd.DataFrame(classification_df, index=[0]), pd.DataFrame(advdetection_df, index=[0]), ) def run_demo(dataset, attacker, text=None, label=None): try: data = { "dataset": dataset, "attacker": attacker, "text": text, "label": label, } response = requests.post('https://rpddemo.pagekite.me/api/generate_adversarial_example', json=data) result = response.json() print(response.json()) return ( result["text"], result["label"], result["restored_text"], result["result_label"], result["perturbed_text"], result["text_diff"], result["perturbed_diff"], result["restored_diff"], result["output"], pd.DataFrame(result["classification_df"]), pd.DataFrame(result["advdetection_df"]), result["message"] ) except Exception as e: print(e) return generate_adversarial_example(dataset, attacker, text, label) def check_gpu(): try: response = requests.post('https://rpddemo.pagekite.me/api/generate_adversarial_example', timeout=3) if response.status_code < 500: return 'GPU available' else: return 'GPU not available' except Exception as e: return 'GPU not available' if __name__ == "__main__": try: init() except Exception as e: print(e) print("Failed to initialize the demo. Please try again later.") demo = gr.Blocks() with demo: gr.Markdown("

Detection and Correction based on Word Importance Ranking (DCWIR)

") gr.Markdown("

Clarifications

") gr.Markdown(""" - This demo has no mechanism to ensure the adversarial example will be correctly repaired by Rapid. The repair success rate is actually the performance reported in the paper.The user must know the resulted output for sake of demonstration. - The adversarial example and corrected adversarial example may be unnatural to read, while it is because the attackers usually generate unnatural perturbations. - All the proposed attacks are Black Box attack where the attacker has no access to the model parameters. """) gr.Markdown("

Natural Example Input

") with gr.Group(): with gr.Row(): input_dataset = gr.Radio( choices=["SST2", "IMDB", "MR", "AGNews10K"], value="SST2", label="Select a testing dataset and an adversarial attacker to generate an adversarial example.", ) input_attacker = gr.Radio( choices=["BAE", "PWWS", "TextFooler", "DeepWordBug"], value="TextFooler", label="Choose an Adversarial Attacker for generating an adversarial example to attack the model.", ) with gr.Group(visible=True): with gr.Row(): input_sentence = gr.Textbox( placeholder="Input a natural example...", label="Alternatively, input a natural example and its original label (from above datasets) to generate an adversarial example.", ) input_label = gr.Textbox( placeholder="Original label, (must be a integer, because we use digits to represent labels in training)", label="Original Label", ) gr.Markdown( "

Default parameters are set according to the main experiment setup in the report.

", ) with gr.Row(): wir_percentage = gr.Textbox( placeholder="Enter percentage from WIR...", label="Percentage from WIR", ) frequency_threshold = gr.Textbox( placeholder="Enter frequency threshold...", label="Frequency Threshold", ) max_candidates = gr.Textbox( placeholder="Enter maximum number of candidates...", label="Maximum Number of Candidates", ) msg_text = gr.Textbox( label="Message", placeholder="This is a message box to show any error messages.", ) button_gen = gr.Button( "Generate an adversarial example to repair using Rapid (GPU: < 1 minute, CPU: 1-10 minutes)", variant="primary", ) gpu_status_text = gr.Textbox( label='GPU status', placeholder="Please click to check", ) button_check = gr.Button( "Check if GPU available", variant="primary" ) button_check.click( fn=check_gpu, inputs=[], outputs=[ gpu_status_text ] ) gr.Markdown("

Generated Adversarial Example and Repaired Adversarial Example

") with gr.Column(): with gr.Group(): with gr.Row(): output_original_example = gr.Textbox(label="Original Example") output_original_label = gr.Textbox(label="Original Label") with gr.Row(): output_adv_example = gr.Textbox(label="Adversarial Example") output_adv_label = gr.Textbox(label="Predicted Label of the Adversarial Example") with gr.Row(): output_repaired_example = gr.Textbox( label="Repaired Adversarial Example by Rapid" ) output_repaired_label = gr.Textbox(label="Predicted Label of the Repaired Adversarial Example") gr.Markdown("

Example Difference (Comparisons)

") gr.Markdown("""

The (+) and (-) in the boxes indicate the added and deleted characters in the adversarial example compared to the original input natural example.

""") ori_text_diff = gr.HighlightedText( label="The Original Natural Example", combine_adjacent=True, show_legend=True, ) adv_text_diff = gr.HighlightedText( label="Character Editions of Adversarial Example Compared to the Natural Example", combine_adjacent=True, show_legend=True, ) restored_text_diff = gr.HighlightedText( label="Character Editions of Repaired Adversarial Example Compared to the Natural Example", combine_adjacent=True, show_legend=True, ) gr.Markdown( "##

The Output of Reactive Perturbation Defocusing

" ) with gr.Row(): with gr.Column(): with gr.Group(): output_is_adv_df = gr.DataFrame( label="Adversarial Example Detection Result" ) gr.Markdown( """ - The is_adversarial field indicates if an adversarial example is detected. - The perturbed_label is the predicted label of the adversarial example. - The confidence field represents the ratio of Inverted samples among the total number of generated candidates. """ ) with gr.Column(): with gr.Group(): output_df = gr.DataFrame( label="Correction Classification Result" ) gr.Markdown( """ - If is_corrected=true, it has been Corrected by DCWIR. - The pred_label field indicates the standard classification result. - The confidence field represents ratio of the dominant class among all Inverted candidates. - The is_correct field indicates whether the predicted label is correct. """ ) # Bind functions to buttons button_gen.click( fn=run_demo, inputs=[input_dataset, input_attacker, input_sentence, input_label], outputs=[ output_original_example, output_original_label, output_repaired_example, output_repaired_label, output_adv_example, ori_text_diff, adv_text_diff, restored_text_diff, output_adv_label, output_df, output_is_adv_df, msg_text ], ) demo.queue(2).launch()