* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download Countering Evolving Threats in Distributed Applications
Deep packet inspection wikipedia , lookup
Information security wikipedia , lookup
Information privacy law wikipedia , lookup
Cyberwarfare wikipedia , lookup
Address space layout randomization wikipedia , lookup
Distributed firewall wikipedia , lookup
Computer and network surveillance wikipedia , lookup
Security-focused operating system wikipedia , lookup
Cross-site scripting wikipedia , lookup
Cyber-security regulation wikipedia , lookup
Cyberattack wikipedia , lookup
Denial-of-service attack wikipedia , lookup
Computer security wikipedia , lookup
Mobile security wikipedia , lookup
Countering Evolving Threats in Distributed Applications: Scientific Principles Saurabh Bagchi The Center for Education and Research in Information Assurance and Security (CERIAS) School of Electrical and Computer Engineering Purdue University Joint work with: Gaspar Howard, Chris Gutierrez, Jeff Avery, Alan Qi (Purdue); Guy Lebanon (Amazon); Donald Steiner (Northrop Grumman) Work Supported By: Northrop Grumman, NSF Slide 1/13 What is Special about Distributed System Security? • Most of our critical infrastructure is built out of careful orchestration of multiple distributed services – Banking, Military mission planning, Power grid, … • Distributed infrastructure means – Many machines, possibly under different admin domains – Many users, external and internal – Dynamic environment where software gets upgraded, new users are added, new machines are added • Attack surface is large and changing – All of the above dynamic factors cause this – Attack may originate from outside or inside Slide 2/13 Three Big Trends in Threats Against Distributed Systems 1. Attack at the point of least resistance – Find a vulnerable outward-facing service, OR – Initiate an insider attack 2. Exploit zero-day vulnerabilities in any constituent service – Thriving black market in zero-day vulnerabilities – Tweak existing attack vectors to bypass rigid defense systems 3. Set up a covert channel for leaking sensitive information – Relevant for systems with highly sensitive but low volume data – Timing channels, storage channels Slide 3/13 Current Approaches against These Three Threat Vectors 1. Attack at the point of least resistance – Create an ever more rigid perimeter – Improve the IDS alerting mechanisms, built alert correlation 2. Exploit zero-day vulnerabilities in any constituent service – Hope white hats (vendors, open source devs) find these before the black hats – Some impactful work in detecting metamorphic malware 3. Set up a covert channel for leaking sensitive information – Only ad-hoc techniques leading to an arms race – Timing channels: perturb timing of actions indiscriminately – Storage channels: “null out” values of all unused storage elements Slide 4/13 Desired Characteristics of Solutions • Clean slate design approach – Build individual services following secure design principles – Includes randomization, use of type safe programming languages, static vulnerability checking, dynamic taint analysis OR • Bolt security on – Embed secure layer on constituent services, not relying only on an impenetrable perimeter – Use the power of big data – lots of users, lots of machines, lots of workloads – Learn from mistakes, i.e., the attacks that succeed – allow expert security admins to provide input to automated system Slide 5/13 A Glimpse into Our Solution Approaches Slide 6/13 Distributed Inferencing from Individual Sensor Information D1 D5 D2 D6 D4 D3 Slide 7/13 Automatic Generation and Update of IDS Signatures: SQLi • First for SQL injection attacks 4. generalized is created for each cluster, 3. Applies aa clustering to the 1. A Crawls multiple cybersecurity portals 2. Extracts richsignature setpublic of technique features from thesamples, attack using logistic regression modeling to collect giving theattack distinctive samples features for each cluster samples 8 Slide 8/13 Automatic General and Update of Signatures: Phishing • Next for phishing attacks • Phishing specific features are created – Word features determined using word frequency counting – Based on common phishing features, e.g., # links, # image tags – Sentiment analysis for determining words conveying sense of change and urgency that attackers attempt to portray to the user • Parsing phishing emails (corpus from Purdue’s IT organization) input as mbox files Slide 9/13 Phishing: Preliminary Results This cluster includes features such as: "below ,need, dear, update, customer, account, bank" • Each cluster forms a general story about the emails contained within it from which the basis of the attack can be deduced – For example, for cluster 4, the attack is trying to get the user to update information for their banking account. • It is much easier training the user based on the attack signature for clusters, than the mass of individual emails Slide 10/13 Covert Timing Channels • Designed a covert network timing channel imitating long range dependent (LRD) legitimate traffic – Can be hidden in the Web traffic, the most observed traffic on Internet today – Statistically indistinguishable from real traffic – Evades the best available detection methods. • Data Rate: 2 – 6 bits/second • Decoding Error: 3% – 6 % • Solution approach – Look for autocorrelation function values – Look for Hurst value that characterizes LRD traffic Slide 11/13 Take Aways • Distributed applications need to be protected • Three emerging trends 1. Attack at the point of least resistance 2. Exploit zero-day vulnerabilities in any constituent service 3. Set up a covert channel for leaking sensitive information • Lessons in solving these trends – If clean slate design is possible for some services, use a comprehensive set of secure design principles: randomization, use of type safe programming languages, static vulnerability checking, dynamic taint analysis – If security needs to be bolted on, look at internal security, not just perimeter security – Big data advances can enable learning from large volumes of existing data to extrapolate to new attack types Slide 12/13 Presentation available at: Dependable Computing Systems Lab (DCSL) web site engineering.purdue.edu/dcsl Slide 13/13