By Christopher Moran, Nicoara Talpes 1
Solution is addressed to VMs that are web servers Web servers should not have confidential information anyway "A complete firewall solution can be created in the cloud by utilizing Amazon EC2’s default deny-all mode which automatically denies all inbound traffic unless the customer explicitly opens an EC2 port. “ – default for instances, protects confidentiality ◦ Meets security outlined in Health Insurance Portability and Accountability Act of
How does the attacker know when the victim launches new instances? Assumes that ‘most people’ use small VMs, but those represent only 21% A web service will not go for small instances if it has decent traffic 3
Co-residence reverse engineering is highly dependent on EC2’s architecture, thus not applicable for other providers Microsoft Azure does not have VM capability Side-channels not useful for inferring encrypted messages Use lab-conditions for keystroke timing tests, instead of the cloud 4
Unrealistic advantages of the testbed: idle machine, no core switching VM cross channel leakage could be explained more clearly, use examples of how an attacker might use the leakage ◦ Live example on the cloud ◦ Explanation of how a hacker might use technique Inefficient covert-channel method : 0.2 bits/sec amounts to 720 bits/hour Unrealistic: did not determine co-residence using more than 2 instances on a CPU 5
Load-based co-residence test will not function for VMs that are not hosting web services: there is a fraction of instances unreachable D.O.S attack impractical on a large share of victim’s instances since coverage is around 8.4 % (at paper’s budget) 6
Paper did not actually do any data theft to prove it is possible. No data for co-residence success rates for victims that have servers up for longer than two days before the attack Extrapolation from covert channel communication between 2 VMs to side-channel model( attacker- victim) is not explained. 7
No cost projection for achieving any results in the paper Experiment should have been designed between two parties, the attackers not knowing the victim’s launch schedules, outside public information Instances cannot be placed with extra-large VMs 8
Simple solutions could be implemented by Amazon: breaking parallel placement by assigning more random IPs to new instances Or remap VMs periodically The paper does not justify the solution of isolating users to different hardware ◦ The point of cloud computing is sharing resources on the same machine and on-demand scaling. ◦ Pushes people to use larger instances when unnecessary 9
Assuming that data on the servers is confidential, is there value gained from the techniques used? Don’t know data hosted on VM, guesstimate machine’s use based on site and CPU activity, will not know specifics about system, intelligence learned is high level ◦ Could have poorly implemented system that overuses CPU for amount of traffic Other keystroke timing attacks known before, this did not require co-residency ◦ Technique relies heavily on knowledge of a person’s typing style ◦ Need to know when they are typing sensitive information 10