Category: systems

NVidia Tiny Linux Kernel and TrustZone

The NVidia Tiny Linux Kernel (TLK), is 23K lines of BSD licensed code, which supports multi-threading, IPC and thread scheduling and implements TrustZone features of a Trusted Execution Environment (TEE). It is based on the Little Kernel for embedded devices.

The TEE is an isolated environment that runs in parallel with an operating system, providing security. It is more secure than an OS and offers a higher level of functionality than a SE, using a hybrid approach that utilizes both hardware and software to protect data. Trusted applications running in a TEE have access to the full power of a device’s main processor and memory, while hardware isolation protects these from user installed apps running in a main operating system. Software and cryptographic isolation inside the TEE protect the trusted applications contained within from each other.

TrustZone was developed by Trusted Foundations Software which was acquired by Gemalto. Giesecke & Devrient developed a rival implementation named Mobicore. In April 2012 ARM, Gemalto and Giesecke & Devrient combined their TrustZone portfolios into a joint venture Trustonic, which was the first to qualify a GlobalPlatform-compliant TEE product in 2013.

A comparison with other hardware based security technologies is found here. Whereas a TPM is exclusively for security functions and does not have access to the CPU,  the TEE does have such access.

Attacks against TrustZone on Android are described in this blackhat talk. With a TEE exploit,  “avc_has_perm” can be modified to bypass SELinux for Android. By the way, Access Vectors in SELinux are described in this wonderful link. “avc_has_perm” is a function to check the AccessVectors allows permission.

Embedded Neural Nets

A key problem for embedded neural networks is reduction of size and power consumption.

The hardware on which the neural net runs on can be a dedicated chip, an FPGA, a GPU or a CPU. Each of these consumes about 10x the power of the previous choice. But in terms of upfront cost, the dedicated chip costs the highest, the CPU the lowest. An NVidia whitepaper compares GPU with CPU on speed and power consumption. (It discusses key  neural networks like AlexNet. The AlexNet was a breakthrough in 2012 showing a neural network to be superior to other image recognition approaches by a wide margin).

Reducing the size of the neural network also reduces its power consumption. For NN size reduction, pruning of the weak connections in the net was proposed in “Learning both Weights and Connections for Efficient Neural Networks” by Song Han and team at NVidia and Stanford. This achieved a roughly 10x reduction in network size without loss of accuracy. Further work in “Deep Compression” achieved a 35x reduction.

Today I attended a talk on SqueezeNet by Forrest Iandola. His team at Berkeley modified (squeezed) the original architecture, then applied the Deep Compression technique above to achieve a 461x size reduction over the original, to 0.5Mb. This makes it feasible for mobile applications. This paper also references the V.Badrinarayan’s work on SegNet – a different NN architecture, discussed in a talk earlier this year.

The Nervana acquisition by Intel earlier this year was for a low power GPU like SOC chip with very high memory bandwidth.

Spark, Storm, Ayasdi, Hadoop

The huge amount of data that IOT systems will generate will call for analyses of different types. A brief review of some systems and what they are used for.

Apache Spark: Uses distributed memory abstractions for primarily in-memory processing. Built with Scala. Useful for finding data clusters and for detecting statistical anomalies by looking for distance from the cluster. Comes with a machine learning system on top. Does not come with its own file system (use nfs/hdfs). Useful for complex processing, where state needs to be maintained in the event stream for correlations. Described as ‘batch processing with micro-streaming analysis’, but looks headed to cover streaming analyses as well.

Apache Storm: Real-time Streaming data analysis. Developed at Twitter, written in Clojure. Unlike Hadoop which has two layers (map, reduce), Storm can have N layers and a flexible topology consisting of Spouts (data source units) and Bolts (data processing units). Storm has been superceded by Heron in terms of performance. IBM Streams is a commercial offering also for stream processing..

Ayasdi: Topological data processing allows one to discover what interesting features of the data are, without knowing what to look for in advance. This is in contrast to most systems where one needs to know what one is looking for. Claims insight discovery.

Hadoop: Used for batch processing of a large amounts of data, using map/reduce primitives. Comes with HDFS. Cloudera (and others) have made significant improvements to it with an interactive SQL interface and usability improvements for BI (Impala).

InfluxDB: Time-series db for events and metrics. Optimized for writes and claims to scale to IOT workloads.

ZooKeeper: A coordination service for distributed applications.

Amazon S2N and OpenSSL

In the last few years a number of OpenSSL vulnerabilities have come to light.  Heartbleed was a critical one which was exploited in the field. It basically allowed a client to send a malicious heartbeat to the server and get back chunks of server memory – which can contain passwords. It was estimated that two thirds of the servers in the world had the vulnerability. The fix was to upgrade OpenSSL, revoke existing server certs and request new SSL server certs.

Heartbleed previously triggered OpenBSD to fork OpenSSL to LibreSSL and Google to fork OpenSSL to BoringSSL.

Amazon S2N is a TLS/SSL implementation that is 6000 lines of code – so it is small, compact, fast and its correctness can be more easily verified. It uses only crypto functions from openssl and reimplements the SSL layer. This is a healthy direction for IOT and for certification of SSL, for example FIPS. S2N is short for Signal to Noise.

A timing attack was recently identified against it and has since been mitigated.

Note that two factor auth solutions would actually solve the problem presented by Heartbleed. There are several solutions in this area – Authy, Clef, Google Authenticator, Duo, Okta, Oracle Authenticator, ..

Docker Container Security

A block diagram of docker is below and a description of docker daemon is here. The docker client commands talk to the docker-daemon to start one of the containers in the docker registry, or to start a process described in the command line as a new docker container. Docker provides a simple interface to linux container technology which is a lightweight VM.


A few problems with this. Who has access to the docker-daemon to control the containers ? How is integrity of the containers ensured ? How is the host protected from the code running in the containers ?

Docker recently announced a few security features in Nov DockerCon

  • to lock down the container in a registry with the container image signed with a key from hardware device Yubikey; see here for a description of original issue where image checksums were not verified by docker daemon
  • to scan the official container images for vulnerabilities
  • to run containers with a userlevel namespace instead of one that allows root access to the host. This protects the host OS as explained here. The userlevel namespace feature has been available in LXC for over an year, but not in docker.

For access control to the docker daemon there is activity with a design doc here.

Twistlock is a container security and monitoring tool that attempts a comprehensive approach – access control to the containers, runtime scanning of files for malware signatures, vulnerability scanning, looking at network packets, so on. A recent meetup on Dec 1 discussed this product. It features integration with Kerberos and LDAP.

In terms of the kernel,  processes from all containers share the same kernel, the same networking layer. So what’s the level of isolation provided to container processes. This depends on vulnerabilities in the processes themselves – how many ports are open, whether injection attacks are possible etc. If two containers are running processes and a process from the one attacks a process from another – for example memory scraping, then Twistlock can detect it only if it can identify the offending process as malware using signature matching.

A Dockerfile is used to specify a container image using commands to spec the base os, rpms, utilities and scripts. USER specifies the userid under which the following RUN, CMD or ENTRYPOINT instruction run. EXPOSE specs a port to be opened for external access. A docker image is built from the dockerfile and contains the actual bits needed for the container to run. The image can be loaded directly or pushed to a docker registry from  which it can be pulled to clients. 

“Computer Detective in the Cloud”

Although light on details, this is an application of AI for securing against credit card fraud in real time using cloud computing.

AI has been in the news a few times this month – Google (TensorFlow), Facebook (new milestones in AI), Microsoft releasing Cortana (Nadella welcomes our AI overlords) and mention of an AI spring from IBM and Salesforce.

Machine learning has also been applied to spam detection, intrusion detection, malicious file detection, malicious url detection, insurance claims leakage detection, activity/behaviour based authentication, threat detection and data loss prevention.

Worth noting that these successes are typically in narrow domains with narrow variations of what is being detected. Intrusion detection is a fairly hard problem for machine learning because the number of variations of attacks is high. As someone said, we’ll be using signatures for a long time.

The previous burst of activity around neural networks in the late 80’s and early 90’s had subsided around the same time as the rise of the internet in the mid to late 90’s. Around 2009, as GPU’s made parallel processing more mainstream, there was a resurgence in activity – deeper, multilayer, networks looking at overlapping regions of images (similar to wavelets) lead to convolutional neural networks being developed. These have had successes in image and voice recognition. A few resources – GPU gems for general purpose computing, visualizing convolutional netscaffe deep learning framework.

Kafka Security

Kafka is a system for continuous, high throughput messaging of event data, such as logs, to enable near real-time analytics. It is structured as a distributed message broker with incoming-event producers sending messages to topics and outgoing-event consumers.  Motivations behind its development include decoupling producers and consumers from each other for flexibility, reducing time to process events and increasing throughput. Couple analogies to think of it are a sender using sendmail to send an email to an email address (topic);  or a message “router” which decides the destination for a particular message – except Kafka persists the messages until the consumer is ready for them. It is an intermediary in the log processing pipeline – there is no processing of data itself on Kafka – there are no reads for instance. In contrast to JMS, one can send batch messages to Kafka and individual messages do not have to be acknowledged.

A design thesis of Kafka is that sequential (contiguous) disk access is very fast and can be even faster than random memory access. It uses zero copy, and uses a binary protocol over TCP, not HTTP.  A quote from design link – “This combination of pagecache and sendfile means that on a Kafka cluster where the consumers are mostly caught up you will see no read activity on the disks whatsoever as they will be serving data entirely from cache”.  This along with the distributed design makes it faster than competing pub-sub systems.

A proposal for adding security to it has been underway, for enterprise use, to control who can publish and subscribe to topics – . A talk on Kafka security by HortonWorks on integrating Kerberos authentication, SSL encryption with Kafka was given at a recent meetup. The slides are at –

Of interest was an incident where the SSL patch caused the cluster to become unstable and increase latencies on a production cluster. The issue was debugged using profiling. Although SSL did increase latencies, this specific issue was narrowed to a bug unrelated to SSL in the same patch which had to do with zero copy.