SESSION ID: TECH-W01 How DHS is Doing Cybersecurity with Content Filtering Tom Ruoff Department of Homeland Security National Protection and Programs Directorate Office of Cybersecurity and Communication/Chief Technology Office
DHS & Content Filtering Bottom Line Up Front Q1. Why is DHS is working on this? A1. Because current signature and detonation approaches are not sufficient to allow control of cyber attacks. Q2. What is better? A2. Content Filtering. Test results indicate emist 3.0.3 is capable of blocking zero day malware at about a 99.5% rate. Q2. What does DHS want to accomplish? A3. Strategically improve cybersecurity. Tactically - stimulate both sides of the supply-demand equation to significantly enable and enhance cybersecurity posture for Federal Executive Branch Departments and Agencies as well as critical infrastructure owners and operators Information Technology systems through use of commercially available technology acquired at market driven cost. DHS wants to facilitate cybersecurity culture change to move time scale from months to milliseconds 2
DHS & Content Filtering What You Get Out of This Talk Agenda 1. Technical understanding of what content filtering is 2. How well it work in neutering malware test results 3. What DHS is doing with this cool stuff to protect itself 4. What are our next steps 5. What can you do with this knowledge 6. Motivation to use this approach to secure your enterprise 3
DHS & Content Filtering WHAT IS CONTENT FILTERING? Defining the terms 4
What is Content Filtering? A filtering technology based on a robust understanding of the syntactic structure and semantic meaning of the file type or protocol being filtered to pass known/validated good content Uses a bit/byte level understanding of the file compare to RFC Decomposed objects into base elements of file type/object protocol specification and then reassembles a clean version that excludes non-essential components Requires access to the file type/protocol specification (RFC) and/or extensive reverse engineering Specs frequently don t match reality so sometimes the decomposition process fails since the object does not de-compose per the specification; a Word doc is sometimes not a Word document per the Word RFC.or a Word document masquerades as a PowerPoint Not signature based Resulting file usually very close to original with minimal damage/changes 5
World of Malware Where Content Filter Fits In Two types of Malware attacks (1 of 2) 1. Syntactic The attacker sends incorrect, malformed, or unexpected data to the system in order to execute an exploit. Within syntactic based attacks there are two main variants: a. Non-compliance with Specification In this attack, the data does not comply with the file format/protocol specification and the software processing that data does not properly handle it leading to a program crash and possible exploit. b. Compliance with Specification In this attack, the data complies with the specification, but an incorrect assumption or decision by the developer on how to implement the specification leads to potential program crash and exploit. For example, suppose a program processes a length delimited file and the specification says that a data field is 128 characters but developer knew that by convention (e.g. common use) that only 16 characters were used so he hardcoded an array to be 16 characters long. If an attacker sent a specification compliant data field with 128 characters of data instead of 16 characters it could lead to a buffer overflow and possible exploit. 6
World of Malware Where Content Filter Fits in Two types of Malware attacks (2 of 2) 2. Semantic The attacker sends structurally correct but logically incorrect data to the system to cause the device to operate outside of its design parameters (e.g. tell a generator to operate 20K RPM above its design tolerance of 5K RPM). 7
So Why Does Content Filtering Work? Most malware very fragile, format conversion changes to the file can break it (render operationally useless) Malware likes to misrepresent itself E.g. a JPEG claiming to be TIFF Malware exploits defects in parsing, usually by providing a structurally wrong or logically incorrect file Malware developers like to hide in the portions of files used for metadata storage, at the end of the file, between segments/markers in a file, and via steganographic techniques in the payload of files (e.g. image data) 8
Content Filtering: Deep Content Inspection & Sanitization ASSUMPTIONS 1. Detecting malware is really hard so don t try 2. Malware is fragile so extracting content and re-assembling objects neuters almost all attacks 3. Exploding the malware is a good start to observe malicious behavior but not entirely effective 4. Active content within object protocol (Excel formulas) are benign the rest is assumed malicious 5. There is a user impact (like rendering URLs inactive) and need to be part of policy settings 6. If the object is not definable (Syntactic attack - kind of a Word 2007 ) then policy can either drop file or pass 9
Content Filtering Methods Deep Content Inspection and Sanitization Verifies file complies with specification, then writes out known good content Format Conversion Converts a file to another related format before converting back to the original file format (e.g. PDF to PS to PDF) File Flattening Converts file to another similar but usually less complex format that doesn t have the data attack risks of the original (e.g. PPT to series of JPG files) Canonicalization Convert contents from specialized form into normalized/raw form (e.g. audio files into PCM) 10
Typical Content Filtering Process Typical Office Document Text Dirty Word Search Based on a Dirty and Clean word list <Image> </Image> Images are inspected for format and sanitized for embedded information or malware <Excel> </Excel> Embedded objects are inspected up to a configurable level deep, usually 1 Macro removal filter <Macro> </Macro> Virus Cleaning 11
How Does it Work: MS Office (1 of 2) Microsoft Office Filters (97-2010), Word (.doc/.docx), Excel (.xls/.xlsx), PowerPoint (.ppt/.pptx) - Processing Steps 1. Validate file type compiles with official specification from Microsoft (2003 and below) or from Microsoft and the ISO for (2007+) 2. Recursively process MS Office into constituent parts 3. Perform text extraction for dirty word analysis 12
How Does it Work: MS Office (2 of 2) Microsoft Office Filters (97-2010), Word (.doc/.docx), Excel (.xls/.xlsx), PowerPoint (.ppt/.pptx) - Processing Steps continued 4. Send all non-ms Office components that are supported to other filters. If file type not supported then either fail the MS Office file or remove that object from the MS Office* 5. Non-MS Office components are filtered by their respective filters and if possible reinserted back into the parent MS office document 13
How Does it Work: Imagery JPEG (.jpg,.jpeg), Windows Bitmap (.bmp/.dib), Windows Metafile (.wmf), Windows Enhanced Metafile (.emf), Graphics Interchange Format (.gif), Portable Network Graphics (.png), Tagged Image File Format (.tiff) Processing Steps: 1. Validate file type compiles with official specification 2. Validate and/or remove metadata 3. Send metadata for dirty word analysis 4. Zeroize the least significant bits of the image data* 5. Rebuild and recompress image * Does not apply to WMF/EMF files 14
How Does it Work: Compressed Files PKzip (.zip), UNIX tar (.tar), GNU zip (.gz), BZip2 (.bz2) Steps: 1. Validate file type compiles with official specification 2. Check excessive levels of embedding (zip/tar) 3. Extract directory structure data 4. Extract all the files and throw away the container 5. Filter files 6. Rebuild container by reinserting filtered files. Failed files are replaced with zero byte files 15
How Does it Work: Text Text files (.txt/.csv/.log) Support 7 bit/8 bit ASCII and Unicode UTF-8 - Steps 1. Validate the file is non-executable textual data 2. Apply Regular Expressions to data (usually to neuter URLs) 3. Apply Dirty Word Filter to textual by rotating through a series of commonly used Code Pages (e.g. character encodings) 16
How Does it Work: PDF Adobe Portable Document Format (PDF) - Processing Steps 1. Validate file type compiles with official specification 2. Perform text extraction for Dirty Word Analysis 3. Convert PDF to Postscript (PS) then back to PDF 4. Validate that encrypted and JavaScript content were removed 17
Content Filtering Lab Test Results Methodology for determining emist s effectiveness at neutralizing malware and determining false positive rates: 1. Collect presumed good and malicious test data. 2. Verify the malicious data using established test bed. 3. Configure emist v3.0.3 with the appropriate policies, network configuration, etc. 4. Process files through emist v3.0.3. 5. Record output results (e.g., passed, modified, rejected) for each file, per file type. 6. Evaluate malicious test set output files for malicious content using established test bed. 7. Analyze results and calculate 95% confidence-level ranges. 18
How Well Does Content Filtering Work Lab Results File Type Block/Cleansing Rate (479 Policy) Block/Cleansing Rate (Basic Policy) Doc 95.28% ± 2.02% 98.63% ± 1.56% Ppt 80.48% ± 24.76% //99% 71.92% ± 33.67% /99% Pdf 99.80% ± 0.16% 99.87% ± 0.18% Xls 96.62% ± 1.33%//98% 98.06% ± 1.43%//98% Gif 98.22% ± 2.50% //100% 96.56% ± 4.78% //100% Jpg 2.91% ± 1.33% 2.88% ± 1.86% Rtf N/A//99.8% N/A//99.8% At 95% Confidence Factor 19
How Well Does Content Filtering Work Lab Testing False Positive Rate File Type False Positive Rate (479 Policy) False Positive Rate (Basic Policy) doc 4.28% ± 0.79% 4.27% ± 1.12 ppt 5.36% ± 1.53% 5.68% ± 2.21% xls 8.26% ± 2.94% 8.73% ± 4.23% docx 5.03% ± 0.50% 44.55% ± 1.62% pptx 15.39% ± 1.10% 25.81% ± 1.89% xlsx 16.73% ± 2.37% 19.16% ± 3.52% pdf 1.49% ± 0.20% 3.39% ± 0.43% gif 1.73% ± 0.58% 1.82% ± 0.84% tiff 1.32% ± 0.32% 1.36% ± 0.46% jpg 1.45% ± 0.31% 1.36% ± 0.42% png 1.66% ± 0.29% 1.83% ± 0.42% bmp 1.88% ± 0.53% 2.03% ± 0.78% wmf 1.25% ± 0.56% 1.31% ± 0.81% emf 1.35% ± 0.42% 1.28% ± 0.57% 95% Confidence Factor 20
Review of Lab Testing Results from testing indicate emist 3.0.3 appears to be capable of blocking zero day malware at about a 99.5% rate Pass rate is 98.5%, can be improved by tailoring dirty word list OR If object is not defined then send to secondary inspection process since this means the object may be malicious take a systems approach 21
DHS Operational Testing of emist 3.0.3 We will put emist 3.0.3 in our operational network (LAN A) to assess operational malicious content kill rate Test results forthcoming: we ran into operational issues so test results need to be verified before public release 22
emist Mail Content Filtering Combined with Behaviorbased Tools DHS SOC OneNet DC2 LAN-A Internet OneNet Hub Transport Server @dhs.gov Email Server MS Outlook Client Main Inbox Current @dhs.gov email path 23
emist Mail Content Filtering Combined with Behaviorbased Tools DHS SOC OneNet DC2 LAN-A Internet OneNet Hub Transport Server @dhs.gov Email Server emist Email Server emist Pilot adds Endpoint Protection ()- equipped laptops, an server, and the emist Mail Content Filtering tool CS&C Participants equipped Laptops 24
emist Mail Content Filtering Combined with Behaviorbased Tools DHS SOC OneNet DC2 LAN-A Internet OneNet Hub Transport Server @dhs.gov Email Server Email traffic entering dhs.gov is replicated and goes to both primary Outlook server and emist emist emist Email Server CS&C Participants equipped Laptops 25
emist Mail Content Filtering Combined with Behaviorbased Tools emist extracts embedded attachments in emails and cleans them emist Emails are reconstructed with their now-cleansed attachments re-inserted 26
emist Mail Content Filtering Combined with Behaviorbased Tools Pilot participants with laptops have Outlook Clients connect to 2 inboxes OneNet Allows tools to detect Hub malicious Transport behavior from files originating from Server either email inbox Internet DHS SOC OneNet DC2 LAN-A @dhs.gov Email Server MS Outlook Client Main Inbox emist emist Email Server Test Inbox CS&C Participants -equipped Laptops 27
emist Mail Content Filtering Combined with Behaviorbased Tools on laptop monitors for and alerts on suspicious behaviors, including reference to files that are source of suspect behaviors Internet DHS SOC OneNet DC2 LAN-A OneNet Hub Transport Server @dhs.gov Email Server MS Outlook Client Main Inbox emist emist Email Server Test Inbox CS&C Participants -equipped Laptops 28
emist Mail Content Filtering Combined with Behaviorbased Tools Data aggregated by server now supports multiple cybersecurity activities -detected behaviors from laptops 29
emist Mail Content Filtering Combined with Behaviorbased Tools -detected behaviors from laptops Malicious items successfully blocked by emist/ missed by current mechanisms 30
emist Mail Content Filtering Combined with Behaviorbased Tools -detected behaviors.gov emails -detected behaviors emist test emails Malicious items not blocked by emist candidates for tuning, signature development, or heuristics 31
DHS Use of Content Filtering What DHS is doing with content filtering to promote its use? We put emist 3.0.3 and follow-on commercial in our operational network (LAN A) to assess operational malicious content kill rate slide show Will use evidence to justify and encourage procurement of commercial content filtering products Partnering with vendors to advance state of art for email and web content filtering 32
DHS Use of Content Filtering What is DHS Doing next with content filtering? Programming next set of commercial product tests and operational demonstrations of kill rate email and web Planning next set of operational tests using a TBD commercial product to perform content filtering on DHS LAN A email Focus will be on sanitization rate, usability and availability Using evidence to justify and encourage procurement of commercial content filtering products Partnering with vendors to advance state of art for email and web content filtering 33
What Can YOU Do with this Knowledge? 1. Research content filtering technology become smarter on pass known good approach 2. Become familiar with current commercial state of art 3. Go get some and protect your networks!!! 4. Demand vendors improve offerings the demand side of supply/demand 5. Developers: Go make better commercial offerings to advance state of art and lower cost through competition 34
Parting Words - Motivation 1. This approach works 98% zero day kill rate 2. It is not monetarily costly, sort of depends 3. This approach impacts user experience (based upon policy to block/pass undefinable objects) this is a good thing as it resets expectations for cost of security 4. Really drives bad guys cost up makes their job harder so maybe we are being strategically impactful 5. Soooo, go get some..market research! 35