$(document).ready(function(){ $('#jailbreak-intro').on('click', 'span', function(e) { e.preventDefault(); if (!$(this).hasClass('selected')) { $('#jailbreak-intro-label > span').removeClass('selected'); $(this).addClass('selected'); let cur_img = $(this).text(); var diagram = document.getElementById("jailbreak-intro-img"); diagram.src = "images/metrics/intro-"+cur_img+".png"; } }); $('#refusal-loss-formula-list').on('click', 'a', function(e) { e.preventDefault(); if (!$(this).hasClass('selected')) { $('.formula').hide(200); $('.formula-list > a').removeClass('selected'); $(this).addClass('selected'); var target = $(this).attr('href'); $(target).show(200); } }); var diagram = document.getElementById("reliability-diagram"), jailbreak_results; $.getJSON("demo_results/jailbreak_defense_results.jsonl", function(data) {jailbreak_results = data;}); $('input[type=radio][name=models]').change(function() {diagram.src = updateFigurePath();}); $('input[type=radio][name=defense]').change(function() { if(this.value == 'ppl') { $("#ppl-slider").slider('enable'); $("#gradient-norm-slider").slider('disable'); }else if (this.value== 'gradient_cuff'){ $("#ppl-slider").slider('disable'); $("#gradient-norm-slider").slider('enable'); }else{ $("#ppl-slider").slider('disable'); $("#gradient-norm-slider").slider('disable'); } diagram.src = updateFigurePath(); }); var bin_handle = $("#gradient-norm-threshold"); $( "#gradient-norm-slider" ).slider({ step: 50, min: 0, max: 800, value: 100, create: function() {bin_handle.text($(this).slider("value"));}, slide: function(event, ui) {bin_handle.text(ui.value);diagram.src = updateFigurePath(ui.value);}, }); var ts_handle = $("#ppl-threshold"); $("#ppl-slider").slider({ step: 20, min: 0, max: 200, value: 100, disabled: true, create: function() {ts_handle.text($(this).slider("value"));}, slide: function( event, ui ) {ts_handle.text(ui.value);diagram.src = updateFigurePath($('#gradient-norm-slider').slider("option", "value"),ui.value);}, }); function updateFigurePath(gradient_norm_threshold=$('#gradient-norm-slider').slider("option", "value"),ppl_threshold=$('#ppl-slider').slider("option", "value")) { var cur_model = $("input[type='radio'][name='models']:checked").val(), cur_defense = $("input[type='radio'][name='defense']:checked").val(); var asr_board = document.getElementById("asr-value"); var brr_board = document.getElementById("brr-value"); if (cur_defense === "ppl") { var asr_gcg=jailbreak_results[cur_defense][cur_model]["adv_harmful_behavior"][ppl_threshold.toString()]["tpr"]; var asr_autodan=jailbreak_results[cur_defense][cur_model]["autodan"][ppl_threshold.toString()]["tpr"]; var asr_pair=jailbreak_results[cur_defense][cur_model]["pair"][ppl_threshold.toString()]["tpr"]; var asr_tap=jailbreak_results[cur_defense][cur_model]["tap"][ppl_threshold.toString()]["tpr"]; var asr_base64=jailbreak_results[cur_defense][cur_model]["base64"][ppl_threshold.toString()]["tpr"]; var asr_lrl_de=jailbreak_results[cur_defense][cur_model]["lrl_de"][ppl_threshold.toString()]["tpr"]; var asr_lrl_fr=jailbreak_results[cur_defense][cur_model]["lrl_fr"][ppl_threshold.toString()]["tpr"]; var asr_lrl_zh_CN=jailbreak_results[cur_defense][cur_model]["lrl_zh-CN"][ppl_threshold.toString()]["tpr"]; var asr_lrl_sv=jailbreak_results[cur_defense][cur_model]["lrl_sv"][ppl_threshold.toString()]["tpr"]; var asr=(asr_gcg+asr_autodan+asr_pair+asr_tap+asr_base64+(asr_lrl_de+asr_lrl_fr+asr_lrl_zh_CN+asr_lrl_sv)/4)/6 asr_board.innerText = parseFloat(asr).toFixed(3); brr_board.innerText = parseFloat(jailbreak_results[cur_defense][cur_model]["pair"][ppl_threshold.toString()]["fpr"]).toFixed(3); return "demo_results/ppl_"+cur_model+"_threshold_"+ppl_threshold.toString()+".png" } else if (cur_defense === "gradient_cuff") { var asr_gcg=jailbreak_results[cur_defense][cur_model]["adv_harmful_behavior"][gradient_norm_threshold.toString()]["tpr"]; var asr_autodan=jailbreak_results[cur_defense][cur_model]["autodan"][gradient_norm_threshold.toString()]["tpr"]; var asr_pair=jailbreak_results[cur_defense][cur_model]["pair"][gradient_norm_threshold.toString()]["tpr"]; var asr_tap=jailbreak_results[cur_defense][cur_model]["tap"][gradient_norm_threshold.toString()]["tpr"]; var asr_base64=jailbreak_results[cur_defense][cur_model]["base64"][gradient_norm_threshold.toString()]["tpr"]; var asr_lrl_de=jailbreak_results[cur_defense][cur_model]["lrl_de"][gradient_norm_threshold.toString()]["tpr"]; var asr_lrl_fr=jailbreak_results[cur_defense][cur_model]["lrl_fr"][gradient_norm_threshold.toString()]["tpr"]; var asr_lrl_zh_CN=jailbreak_results[cur_defense][cur_model]["lrl_zh-CN"][gradient_norm_threshold.toString()]["tpr"]; var asr_lrl_sv=jailbreak_results[cur_defense][cur_model]["lrl_sv"][gradient_norm_threshold.toString()]["tpr"]; var asr=(asr_gcg+asr_autodan+asr_pair+asr_tap+asr_base64+(asr_lrl_de+asr_lrl_fr+asr_lrl_zh_CN+asr_lrl_sv)/4)/6 asr_board.innerText = parseFloat(asr).toFixed(3); brr_board.innerText = parseFloat(jailbreak_results[cur_defense][cur_model]["pair"][gradient_norm_threshold.toString()]["fpr"]).toFixed(3); return "demo_results/gradient_cuff_"+cur_model+"_threshold_"+gradient_norm_threshold.toString()+".png" }else{ var asr_gcg=jailbreak_results[cur_defense][cur_model]["adv_harmful_behavior"]["tpr"]; var asr_autodan=jailbreak_results[cur_defense][cur_model]["autodan"]["tpr"]; var asr_pair=jailbreak_results[cur_defense][cur_model]["pair"]["tpr"]; var asr_tap=jailbreak_results[cur_defense][cur_model]["tap"]["tpr"]; var asr_base64=jailbreak_results[cur_defense][cur_model]["base64"]["tpr"]; var asr_lrl_de=jailbreak_results[cur_defense][cur_model]["lrl_de"]["tpr"]; var asr_lrl_fr=jailbreak_results[cur_defense][cur_model]["lrl_fr"]["tpr"]; var asr_lrl_zh_CN=jailbreak_results[cur_defense][cur_model]["lrl_zh-CN"]["tpr"]; var asr_lrl_sv=jailbreak_results[cur_defense][cur_model]["lrl_sv"]["tpr"]; var asr=(asr_gcg+asr_autodan+asr_pair+asr_tap+asr_base64+(asr_lrl_de+asr_lrl_fr+asr_lrl_zh_CN+asr_lrl_sv)/4)/6 asr_board.innerText = parseFloat(asr).toFixed(3); brr_board.innerText = parseFloat(jailbreak_results[cur_defense][cur_model]["pair"]["fpr"]).toFixed(3); return "demo_results/"+cur_defense+"_"+cur_model+".png" } } });