图像对抗样本的生成（FGSM）

<div class="article_content clearfix csdn-tracking-statistics" id="article_content" style="height: 2436px; overflow: hidden;" data-mod="popu_307" data-pid="blog" data-dsm="post">
                                <link href="https://csdnimg.cn/release/phoenix/template/css/ck_htmledit_views-f57960eb32.css" rel="stylesheet">
                                                    <link href="https://csdnimg.cn/release/phoenix/template/css/ck_htmledit_views-f57960eb32.css" rel="stylesheet">
                        <div class="htmledit_views" id="content_views">
                <p>&nbsp; &nbsp; &nbsp; &nbsp; 快速梯度攻击（Fast Gradient Sign Method，FGSM）由Ian Goodfllow 提出，算法的主要思想是寻找深度学习模型（CNN）的梯度变化最大的方向，按照此方向添加对抗扰动，导致模型进行错误的分类。FGSM以增加对图像分类器损失的方式来对图像添加扰动。通过FGSM构造对抗样本的优势是效率比较高，最终生成的对抗样本会对原图所有像素点都产生一些微弱的视觉暇疵，如下图所示。该攻击属于白盒攻击，需要得到模型网络结构信息。</p>

<p style="text-align:center;"><img width="600" height="242" class="has" alt="" src="https://img-blog.csdnimg.cn/20190306001836115.jpg?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXFpX2Zhbg==,size_16,color_FFFFFF,t_70"></p>

<p>&nbsp; &nbsp; &nbsp; &nbsp;本博客带大家在ImageNet图像上利用FGSM攻击算法生成对抗样本，具体算法实现代码如下：</p>

<pre class="has" name="code"><code class="hljs python"><ol class="hljs-ln hundred"><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="1"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-comment">#!usr/bin/python</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="2"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-comment"># -*- coding: utf-8 -*-</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="3"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-keyword">import</span> os</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="4"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-keyword">from</span> cleverhans.attacks <span class="hljs-keyword">import</span> FastGradientMethod</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="5"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-keyword">import</span> numpy <span class="hljs-keyword">as</span> np</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="6"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-keyword">from</span> PIL <span class="hljs-keyword">import</span> Image</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="7"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-keyword">from</span> scipy.misc <span class="hljs-keyword">import</span> imread</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="8"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-keyword">import</span> tensorflow <span class="hljs-keyword">as</span> tf</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="9"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-keyword">from</span> tensorflow.contrib.slim.nets <span class="hljs-keyword">import</span> inception</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="10"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">slim = tf.contrib.slim</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="11"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">tensorflow_master = <span class="hljs-string">""</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="12"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">checkpoint_path   = <span class="hljs-string">"inception_v3.ckpt所在的路径"</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="13"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">input_dir         = <span class="hljs-string">"合法图像的路径"</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="14"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">output_dir        = <span class="hljs-string">"对抗样本的输出路径"</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="15"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">max_epsilon       = <span class="hljs-number">16.0</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="16"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">image_width       = <span class="hljs-number">299</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="17"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">image_height      = <span class="hljs-number">299</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="18"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">batch_size        = <span class="hljs-number">50</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="19"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-keyword">import</span> sys</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="20"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">sys.path.append(<span class="hljs-string">'cleverhans文件夹所在的路径'</span>)</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="21"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">eps = <span class="hljs-number">2.0</span> * max_epsilon / <span class="hljs-number">255.0</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="22"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">batch_shape = [batch_size, image_height, image_width, <span class="hljs-number">3</span>]</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="23"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">num_classes = <span class="hljs-number">1001</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="24"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"> </div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="25"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">load_images</span><span class="hljs-params">(input_dir, batch_shape)</span>:</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="26"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    images = np.zeros(batch_shape)</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="27"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    filenames = []</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="28"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    idx = <span class="hljs-number">0</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="29"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    batch_size = batch_shape[<span class="hljs-number">0</span>]</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="30"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    <span class="hljs-keyword">for</span> filepath <span class="hljs-keyword">in</span> sorted(tf.gfile.Glob(os.path.join(input_dir, <span class="hljs-string">'*.png'</span>))):</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="31"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        <span class="hljs-keyword">with</span> tf.gfile.Open(filepath, <span class="hljs-string">"rb"</span>) <span class="hljs-keyword">as</span> f:</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="32"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">            images[idx, :, :, :] = imread(f, mode=<span class="hljs-string">'RGB'</span>).astype(np.float)*<span class="hljs-number">2.0</span>/<span class="hljs-number">255.0</span> - <span class="hljs-number">1.0</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="33"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        filenames.append(os.path.basename(filepath))</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="34"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        idx += <span class="hljs-number">1</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="35"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        <span class="hljs-keyword">if</span> idx == batch_size:</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="36"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">            <span class="hljs-keyword">yield</span> filenames, images</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="37"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">            filenames = []</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="38"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">            images = np.zeros(batch_shape)</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="39"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">            idx = <span class="hljs-number">0</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="40"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    <span class="hljs-keyword">if</span> idx &gt; <span class="hljs-number">0</span>:</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="41"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        <span class="hljs-keyword">yield</span> filenames, images</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="42"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"> </div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="43"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"> </div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="44"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">save_images</span><span class="hljs-params">(images, filenames, output_dir)</span>:</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="45"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"> </div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="46"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">  <span class="hljs-keyword">for</span> i, filename <span class="hljs-keyword">in</span> enumerate(filenames):</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="47"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    <span class="hljs-comment"># Images for inception classifier are normalized to be in [-1, 1] interval,</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="48"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    <span class="hljs-comment"># so rescale them back to [0, 1].</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="49"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    <span class="hljs-keyword">with</span> tf.gfile.Open(os.path.join(output_dir, filename), <span class="hljs-string">'w'</span>) <span class="hljs-keyword">as</span> f:</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="50"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">      img = (((images[i, :, :, :] + <span class="hljs-number">1.0</span>) * <span class="hljs-number">0.5</span>) * <span class="hljs-number">255.0</span>).astype(np.uint8)</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="51"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">      Image.fromarray(img).save(f, format=<span class="hljs-string">'PNG'</span>)</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="52"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"> </div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="53"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"> </div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="54"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-class"><span class="hljs-keyword">class</span> <span class="hljs-title">InceptionModel</span><span class="hljs-params">(object)</span>:</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="55"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"> </div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="56"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    <span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">__init__</span><span class="hljs-params">(self, num_classes)</span>:</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="57"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        self.num_classes = num_classes</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="58"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        self.built = <span class="hljs-keyword">False</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="59"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"> </div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="60"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    <span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">__call__</span><span class="hljs-params">(self, x_input)</span>:</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="61"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        <span class="hljs-string">"""Constructs model and return probabilities for given input."""</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="62"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        reuse = <span class="hljs-keyword">True</span> <span class="hljs-keyword">if</span> self.built <span class="hljs-keyword">else</span> <span class="hljs-keyword">None</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="63"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        <span class="hljs-keyword">with</span> slim.arg_scope(inception.inception_v3_arg_scope()):</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="64"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">            _, end_points = inception.inception_v3(</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="65"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">                            x_input, num_classes=self.num_classes, is_training=<span class="hljs-keyword">False</span>,</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="66"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">                            reuse=reuse)</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="67"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        self.built = <span class="hljs-keyword">True</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="68"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        output = end_points[<span class="hljs-string">'Predictions'</span>]</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="69"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        probs = output.op.inputs[<span class="hljs-number">0</span>]</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="70"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        print(output)</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="71"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        <span class="hljs-keyword">return</span> probs</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="72"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"> </div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="73"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-comment"># 加载图片</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="74"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">image_iterator = load_images(input_dir, batch_shape)</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="75"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"> </div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="76"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-comment"># 得到第一个batch的图片</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="77"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">filenames, images = next(image_iterator)</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="78"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-comment"># 日志</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="79"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">tf.logging.set_verbosity(tf.logging.INFO)</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="80"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"> </div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="81"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"><span class="hljs-keyword">with</span> tf.Graph().as_default():</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="82"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    x_input = tf.placeholder(tf.float32, shape=batch_shape)</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="83"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    <span class="hljs-comment"># 实例一个model</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="84"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    model = InceptionModel(num_classes)</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="85"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    <span class="hljs-comment"># 开启一个会话</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="86"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    ses = tf.Session()</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="87"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    <span class="hljs-comment"># 对抗攻击开始</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="88"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    fgsm = FastGradientMethod(model)</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="89"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    x_adv = fgsm.generate(x_input, eps=eps, ord=np.inf, clip_min=<span class="hljs-number">-1.</span>, clip_max=<span class="hljs-number">1.</span>)</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="90"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"> </div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="91"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    <span class="hljs-comment"># fgsm_a是基于L2范数生成的对抗样本</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="92"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    <span class="hljs-comment"># fgsm_a = FastGradientMethod(model)</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="93"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    <span class="hljs-comment"># x_adv = fgsm_a.generate(x_input, ord=2, clip_min=-1., clip_max=1.)</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="94"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line"> </div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="95"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    <span class="hljs-comment"># 恢复inception-v3 模型</span></div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="96"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    saver = tf.train.Saver(slim.get_model_variables())</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="97"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    session_creator = tf.train.ChiefSessionCreator(</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="98"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">                      scaffold=tf.train.Scaffold(saver=saver),</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="99"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">                      checkpoint_filename_with_path=checkpoint_path,</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="100"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">                      master=tensorflow_master)</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="101"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">    <span class="hljs-keyword">for</span> filenames, images <span class="hljs-keyword">in</span> load_images(input_dir, batch_shape):</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="102"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        <span class="hljs-keyword">with</span> tf.train.MonitoredSession(session_creator=session_creator) <span class="hljs-keyword">as</span> sess:</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="103"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">            nontargeted_images = sess.run(x_adv, feed_dict={x_input: images})</div></div></li><li><div class="hljs-ln-numbers"><div class="hljs-ln-line hljs-ln-n" data-line-number="104"></div></div><div class="hljs-ln-code"><div class="hljs-ln-line">        save_images(nontargeted_images, filenames, output_dir)</div></div></li></ol></code><div class="hljs-button {2}" οnclick="hljs.copyCode(event)" data-title="复制"></div></pre>

<p>&nbsp;代码运行之前需要准备一下几个文件：</p>

<ul><li>Inception-v3模型 --&nbsp;<a href="https://github.com/tensorflow/cleverhans/blob/master/examples/nips17_adversarial_competition/dev_toolkit/sample_attacks/download_checkpoints.sh" target="_blank" rel="nofollow">链接</a></li>
    <li>ImageNet合法图像 --&nbsp;&nbsp;<a href="https://github.com/tensorflow/cleverhans/tree/master/examples/nips17_adversarial_competition/dataset" target="_blank" rel="nofollow">链接</a></li>
    <li>cleverhans 库 --&nbsp;<a href="https://github.com/tensorflow/cleverhans/tree/master/cleverhans" target="_blank" rel="nofollow">链接</a></li>
</ul><p>代码中的eps变量的值控制着FGSM的攻击强度，eps值越大生成的对抗样本视觉暇疵越大，视觉效果图如下。</p>

<p style="text-align:center;"><img class="has" alt="" src="https://img-blog.csdnimg.cn/2019030600413230.jpg?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXFpX2Zhbg==,size_16,color_FFFFFF,t_70"></p>

<p>&nbsp;</p>
            </div>
                      </div>