1 // Copyright (c) 2012 The Chromium Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file.
5 #include "chrome/renderer/safe_browsing/phishing_dom_feature_extractor.h"
8 #include "base/compiler_specific.h"
9 #include "base/containers/hash_tables.h"
10 #include "base/logging.h"
11 #include "base/message_loop/message_loop.h"
12 #include "base/metrics/histogram.h"
13 #include "base/strings/string_util.h"
14 #include "base/time/time.h"
15 #include "chrome/renderer/safe_browsing/feature_extractor_clock.h"
16 #include "chrome/renderer/safe_browsing/features.h"
17 #include "content/public/renderer/render_view.h"
18 #include "net/base/registry_controlled_domains/registry_controlled_domain.h"
19 #include "third_party/WebKit/public/platform/WebString.h"
20 #include "third_party/WebKit/public/web/WebElement.h"
21 #include "third_party/WebKit/public/web/WebElementCollection.h"
22 #include "third_party/WebKit/public/web/WebLocalFrame.h"
23 #include "third_party/WebKit/public/web/WebView.h"
25 namespace safe_browsing
{
27 // This time should be short enough that it doesn't noticeably disrupt the
28 // user's interaction with the page.
29 const int PhishingDOMFeatureExtractor::kMaxTimePerChunkMs
= 10;
31 // Experimenting shows that we get a reasonable gain in performance by
32 // increasing this up to around 10, but there's not much benefit in
33 // increasing it past that.
34 const int PhishingDOMFeatureExtractor::kClockCheckGranularity
= 10;
36 // This should be longer than we expect feature extraction to take on any
37 // actual phishing page.
38 const int PhishingDOMFeatureExtractor::kMaxTotalTimeMs
= 500;
40 // Intermediate state used for computing features. See features.h for
41 // descriptions of the DOM features that are computed.
42 struct PhishingDOMFeatureExtractor::PageFeatureState
{
43 // Link related features
45 base::hash_set
<std::string
> external_domains
;
49 // Form related features
55 int action_other_domain
;
58 // Image related features
62 // How many script tags
65 // The time at which we started feature extraction for the current page.
66 base::TimeTicks start_time
;
68 // The number of iterations we've done for the current extraction.
71 explicit PageFeatureState(base::TimeTicks start_time_ticks
)
80 action_other_domain(0),
85 start_time(start_time_ticks
),
88 ~PageFeatureState() {}
92 struct PhishingDOMFeatureExtractor::FrameData
{
93 // This is our reference to document.all, which is an iterator over all
94 // of the elements in the document. It keeps track of our current position.
95 blink::WebElementCollection elements
;
96 // The domain of the document URL, stored here so that we don't need to
97 // recompute it every time it's needed.
101 PhishingDOMFeatureExtractor::PhishingDOMFeatureExtractor(
102 content::RenderView
* render_view
,
103 FeatureExtractorClock
* clock
)
104 : render_view_(render_view
),
106 weak_factory_(this) {
110 PhishingDOMFeatureExtractor::~PhishingDOMFeatureExtractor() {
111 // The RenderView should have called CancelPendingExtraction() before
113 CheckNoPendingExtraction();
116 void PhishingDOMFeatureExtractor::ExtractFeatures(
117 FeatureMap
* features
,
118 const DoneCallback
& done_callback
) {
119 // The RenderView should have called CancelPendingExtraction() before
120 // starting a new extraction, so DCHECK this.
121 CheckNoPendingExtraction();
122 // However, in an opt build, we will go ahead and clean up the pending
123 // extraction so that we can start in a known state.
124 CancelPendingExtraction();
126 features_
= features
;
127 done_callback_
= done_callback
;
129 page_feature_state_
.reset(new PageFeatureState(clock_
->Now()));
130 blink::WebView
* web_view
= render_view_
->GetWebView();
131 if (web_view
&& web_view
->mainFrame()) {
132 cur_document_
= web_view
->mainFrame()->document();
135 base::MessageLoop::current()->PostTask(
137 base::Bind(&PhishingDOMFeatureExtractor::ExtractFeaturesWithTimeout
,
138 weak_factory_
.GetWeakPtr()));
141 void PhishingDOMFeatureExtractor::CancelPendingExtraction() {
142 // Cancel any pending callbacks, and clear our state.
143 weak_factory_
.InvalidateWeakPtrs();
147 void PhishingDOMFeatureExtractor::ExtractFeaturesWithTimeout() {
148 DCHECK(page_feature_state_
.get());
149 ++page_feature_state_
->num_iterations
;
150 base::TimeTicks current_chunk_start_time
= clock_
->Now();
152 if (cur_document_
.isNull()) {
153 // This will only happen if we weren't able to get the document for the
154 // main frame. We'll treat this as an extraction failure.
159 int num_elements
= 0;
160 for (; !cur_document_
.isNull(); cur_document_
= GetNextDocument()) {
161 blink::WebElement cur_element
;
162 if (cur_frame_data_
.get()) {
163 // We're resuming traversal of a frame, so just advance to the next
165 cur_element
= cur_frame_data_
->elements
.nextItem();
166 // When we resume the traversal, the first call to nextItem() potentially
167 // has to walk through the document again from the beginning, if it was
168 // modified between our chunks of work. Log how long this takes, so we
169 // can tell if it's too slow.
170 UMA_HISTOGRAM_TIMES("SBClientPhishing.DOMFeatureResumeTime",
171 clock_
->Now() - current_chunk_start_time
);
173 // We just moved to a new frame, so update our frame state
174 // and advance to the first element.
176 cur_element
= cur_frame_data_
->elements
.firstItem();
179 for (; !cur_element
.isNull();
180 cur_element
= cur_frame_data_
->elements
.nextItem()) {
181 if (cur_element
.hasHTMLTagName("a")) {
182 HandleLink(cur_element
);
183 } else if (cur_element
.hasHTMLTagName("form")) {
184 HandleForm(cur_element
);
185 } else if (cur_element
.hasHTMLTagName("img")) {
186 HandleImage(cur_element
);
187 } else if (cur_element
.hasHTMLTagName("input")) {
188 HandleInput(cur_element
);
189 } else if (cur_element
.hasHTMLTagName("script")) {
190 HandleScript(cur_element
);
193 if (++num_elements
>= kClockCheckGranularity
) {
195 base::TimeTicks now
= clock_
->Now();
196 if (now
- page_feature_state_
->start_time
>=
197 base::TimeDelta::FromMilliseconds(kMaxTotalTimeMs
)) {
198 DLOG(ERROR
) << "Feature extraction took too long, giving up";
199 // We expect this to happen infrequently, so record when it does.
200 UMA_HISTOGRAM_COUNTS("SBClientPhishing.DOMFeatureTimeout", 1);
204 base::TimeDelta chunk_elapsed
= now
- current_chunk_start_time
;
206 base::TimeDelta::FromMilliseconds(kMaxTimePerChunkMs
)) {
207 // The time limit for the current chunk is up, so post a task to
208 // continue extraction.
210 // Record how much time we actually spent on the chunk. If this is
211 // much higher than kMaxTimePerChunkMs, we may need to adjust the
212 // clock granularity.
213 UMA_HISTOGRAM_TIMES("SBClientPhishing.DOMFeatureChunkTime",
215 base::MessageLoop::current()->PostTask(
218 &PhishingDOMFeatureExtractor::ExtractFeaturesWithTimeout
,
219 weak_factory_
.GetWeakPtr()));
222 // Otherwise, continue.
226 // We're done with this frame, recalculate the FrameData when we
227 // advance to the next frame.
228 cur_frame_data_
.reset();
235 void PhishingDOMFeatureExtractor::HandleLink(
236 const blink::WebElement
& element
) {
237 // Count the number of times we link to a different host.
238 if (!element
.hasAttribute("href")) {
239 DVLOG(1) << "Skipping anchor tag with no href";
243 // Retrieve the link and resolve the link in case it's relative.
244 blink::WebURL full_url
= element
.document().completeURL(
245 element
.getAttribute("href"));
248 bool is_external
= IsExternalDomain(full_url
, &domain
);
249 if (domain
.empty()) {
250 DVLOG(1) << "Could not extract domain from link: " << full_url
;
255 ++page_feature_state_
->external_links
;
257 // Record each unique domain that we link to.
258 page_feature_state_
->external_domains
.insert(domain
);
261 // Check how many are https links.
262 if (GURL(full_url
).SchemeIs("https")) {
263 ++page_feature_state_
->secure_links
;
266 ++page_feature_state_
->total_links
;
269 void PhishingDOMFeatureExtractor::HandleForm(
270 const blink::WebElement
& element
) {
271 // Increment the number of forms on this page.
272 ++page_feature_state_
->num_forms
;
274 // Record whether the action points to a different domain.
275 if (!element
.hasAttribute("action")) {
279 blink::WebURL full_url
= element
.document().completeURL(
280 element
.getAttribute("action"));
283 bool is_external
= IsExternalDomain(full_url
, &domain
);
284 if (domain
.empty()) {
285 DVLOG(1) << "Could not extract domain from form action: " << full_url
;
290 ++page_feature_state_
->action_other_domain
;
292 ++page_feature_state_
->total_actions
;
295 void PhishingDOMFeatureExtractor::HandleImage(
296 const blink::WebElement
& element
) {
297 if (!element
.hasAttribute("src")) {
298 DVLOG(1) << "Skipping img tag with no src";
301 // Record whether the image points to a different domain.
302 blink::WebURL full_url
= element
.document().completeURL(
303 element
.getAttribute("src"));
305 bool is_external
= IsExternalDomain(full_url
, &domain
);
306 if (domain
.empty()) {
307 DVLOG(1) << "Could not extract domain from image src: " << full_url
;
312 ++page_feature_state_
->img_other_domain
;
314 ++page_feature_state_
->total_imgs
;
317 void PhishingDOMFeatureExtractor::HandleInput(
318 const blink::WebElement
& element
) {
319 // The HTML spec says that if the type is unspecified, it defaults to text.
320 // In addition, any unrecognized type will be treated as a text input.
322 // Note that we use the attribute value rather than
323 // WebFormControlElement::formControlType() for consistency with the
324 // way the phishing classification model is created.
325 std::string type
= element
.getAttribute("type").utf8();
326 base::StringToLowerASCII(&type
);
327 if (type
== "password") {
328 ++page_feature_state_
->num_pswd_inputs
;
329 } else if (type
== "radio") {
330 ++page_feature_state_
->num_radio_inputs
;
331 } else if (type
== "checkbox") {
332 ++page_feature_state_
->num_check_inputs
;
333 } else if (type
!= "submit" && type
!= "reset" && type
!= "file" &&
334 type
!= "hidden" && type
!= "image" && type
!= "button") {
335 // Note that there are a number of new input types in HTML5 that are not
336 // handled above. For now, we will consider these as text inputs since
337 // they could be used to capture user input.
338 ++page_feature_state_
->num_text_inputs
;
342 void PhishingDOMFeatureExtractor::HandleScript(
343 const blink::WebElement
& element
) {
344 ++page_feature_state_
->num_script_tags
;
347 void PhishingDOMFeatureExtractor::CheckNoPendingExtraction() {
348 DCHECK(done_callback_
.is_null());
349 DCHECK(!cur_frame_data_
.get());
350 DCHECK(cur_document_
.isNull());
351 if (!done_callback_
.is_null() || cur_frame_data_
.get() ||
352 !cur_document_
.isNull()) {
353 LOG(ERROR
) << "Extraction in progress, missing call to "
354 << "CancelPendingExtraction";
358 void PhishingDOMFeatureExtractor::RunCallback(bool success
) {
359 // Record some timing stats that we can use to evaluate feature extraction
360 // performance. These include both successful and failed extractions.
361 DCHECK(page_feature_state_
.get());
362 UMA_HISTOGRAM_COUNTS("SBClientPhishing.DOMFeatureIterations",
363 page_feature_state_
->num_iterations
);
364 UMA_HISTOGRAM_TIMES("SBClientPhishing.DOMFeatureTotalTime",
365 clock_
->Now() - page_feature_state_
->start_time
);
367 DCHECK(!done_callback_
.is_null());
368 done_callback_
.Run(success
);
372 void PhishingDOMFeatureExtractor::Clear() {
374 done_callback_
.Reset();
375 cur_frame_data_
.reset(NULL
);
376 cur_document_
.reset();
379 void PhishingDOMFeatureExtractor::ResetFrameData() {
380 DCHECK(!cur_document_
.isNull());
381 DCHECK(!cur_frame_data_
.get());
383 cur_frame_data_
.reset(new FrameData());
384 cur_frame_data_
->elements
= cur_document_
.all();
385 cur_frame_data_
->domain
=
386 net::registry_controlled_domains::GetDomainAndRegistry(
388 net::registry_controlled_domains::EXCLUDE_PRIVATE_REGISTRIES
);
391 blink::WebDocument
PhishingDOMFeatureExtractor::GetNextDocument() {
392 DCHECK(!cur_document_
.isNull());
393 blink::WebFrame
* frame
= cur_document_
.frame();
394 // Advance to the next frame that contains a document, with no wrapping.
396 for (frame
= frame
->traverseNext(false); frame
;
397 frame
= frame
->traverseNext(false)) {
398 if (!frame
->document().isNull()) {
399 return frame
->document();
403 // Keep track of how often frame traversal got "stuck" due to the
404 // current subdocument getting removed from the frame tree.
405 UMA_HISTOGRAM_COUNTS("SBClientPhishing.DOMFeatureFrameRemoved", 1);
407 return blink::WebDocument();
410 bool PhishingDOMFeatureExtractor::IsExternalDomain(const GURL
& url
,
411 std::string
* domain
) const {
413 DCHECK(cur_frame_data_
.get());
415 if (cur_frame_data_
->domain
.empty()) {
419 // TODO(bryner): Ensure that the url encoding is consistent with the features
421 if (url
.HostIsIPAddress()) {
422 domain
->assign(url
.host());
424 domain
->assign(net::registry_controlled_domains::GetDomainAndRegistry(
425 url
, net::registry_controlled_domains::EXCLUDE_PRIVATE_REGISTRIES
));
428 return !domain
->empty() && *domain
!= cur_frame_data_
->domain
;
431 void PhishingDOMFeatureExtractor::InsertFeatures() {
432 DCHECK(page_feature_state_
.get());
434 if (page_feature_state_
->total_links
> 0) {
435 // Add a feature for the fraction of times the page links to an external
436 // domain vs. an internal domain.
437 double link_freq
= static_cast<double>(
438 page_feature_state_
->external_links
) /
439 page_feature_state_
->total_links
;
440 features_
->AddRealFeature(features::kPageExternalLinksFreq
, link_freq
);
442 // Add a feature for each unique domain that we're linking to
443 for (base::hash_set
<std::string
>::iterator it
=
444 page_feature_state_
->external_domains
.begin();
445 it
!= page_feature_state_
->external_domains
.end(); ++it
) {
446 features_
->AddBooleanFeature(features::kPageLinkDomain
+ *it
);
449 // Fraction of links that use https.
450 double secure_freq
= static_cast<double>(
451 page_feature_state_
->secure_links
) / page_feature_state_
->total_links
;
452 features_
->AddRealFeature(features::kPageSecureLinksFreq
, secure_freq
);
455 // Record whether forms appear and whether various form elements appear.
456 if (page_feature_state_
->num_forms
> 0) {
457 features_
->AddBooleanFeature(features::kPageHasForms
);
459 if (page_feature_state_
->num_text_inputs
> 0) {
460 features_
->AddBooleanFeature(features::kPageHasTextInputs
);
462 if (page_feature_state_
->num_pswd_inputs
> 0) {
463 features_
->AddBooleanFeature(features::kPageHasPswdInputs
);
465 if (page_feature_state_
->num_radio_inputs
> 0) {
466 features_
->AddBooleanFeature(features::kPageHasRadioInputs
);
468 if (page_feature_state_
->num_check_inputs
> 0) {
469 features_
->AddBooleanFeature(features::kPageHasCheckInputs
);
472 // Record fraction of form actions that point to a different domain.
473 if (page_feature_state_
->total_actions
> 0) {
474 double action_freq
= static_cast<double>(
475 page_feature_state_
->action_other_domain
) /
476 page_feature_state_
->total_actions
;
477 features_
->AddRealFeature(features::kPageActionOtherDomainFreq
,
481 // Record how many image src attributes point to a different domain.
482 if (page_feature_state_
->total_imgs
> 0) {
483 double img_freq
= static_cast<double>(
484 page_feature_state_
->img_other_domain
) /
485 page_feature_state_
->total_imgs
;
486 features_
->AddRealFeature(features::kPageImgOtherDomainFreq
, img_freq
);
489 // Record number of script tags (discretized for numerical stability.)
490 if (page_feature_state_
->num_script_tags
> 1) {
491 features_
->AddBooleanFeature(features::kPageNumScriptTagsGTOne
);
492 if (page_feature_state_
->num_script_tags
> 6) {
493 features_
->AddBooleanFeature(features::kPageNumScriptTagsGTSix
);
498 } // namespace safe_browsing