Secure Communication Channel Establishment: TLS 1.3 (over TCP Fast Open) versus QUIC
Secure channel establishment protocols such as Transport Layer Security (TLS) are some of the most important cryptographic protocols, enabling the encryption of Internet traffic. Reducing latency (the number of interactions between parties before encrypted data can be transmitted) in such protocols has become an important design goal to improve user experience. The most important protocols addressing this goal are TLS 1.3, the latest TLS version standardized in 2018 to replace the widely deployed TLS 1.2, and Quick UDP Internet Connections (QUIC), a secure transport protocol from Google that is implemented in the Chrome browser. There have been a number of formal security analyses for TLS 1.3 and QUIC, but their security, when layered with their underlying transport protocols, cannot be easily compared. Our work is the first to thoroughly compare the security and availability properties of these protocols. Toward this goal, we develop novel security models that permit “layered” security analysis. In addition to the standard goals of server authentication and data confidentiality and integrity, we consider the goals of IP spoofing prevention, key exchange packet integrity, secure channel header integrity, and reset authentication, which capture a range of practical threats not usually taken into account by existing security models that focus mainly on the cryptographic cores of the protocols. Equipped with our new models we provide a detailed comparison of three low-latency layered protocols: TLS 1.3 over TCP Fast Open (TFO), QUIC over UDP, and QUIC[TLS] (a new design for QUIC that uses TLS 1.3 key exchange) over UDP. In particular, we show that TFO’s cookie mechanism does provably achieve the security goal of IP spoofing prevention. Additionally, we find several new availability attacks that manipulate the early key exchange packets without being detected by the communicating parties. By including packet-level attacks in our analysis, our results shed light on how the reliability, flow control, and congestion control of the above layered protocols compare, in adversarial settings. We hope that our models will help protocol designers in their future protocol analyses and that our results will help practitioners better understand the advantages and limitations of secure channel establishment protocols.
Cryptanalytic Extraction of Neural Network Models 📺
We argue that the machine learning problem of model extraction is actually a cryptanalytic problem in disguise, and should be studied as such. Given oracle access to a neural network, we introduce a differential attack that can efficiently steal the parameters of the remote model up to floating point precision. Our attack relies on the fact that ReLU neural networks are piecewise linear functions, and thus queries at the critical points reveal information about the model parameters. We evaluate our attack on multiple neural network models and extract models that are 2^20 times more precise and require 100x fewer queries than prior work. For example, we extract a 100,000 parameter neural network trained on the MNIST digit recognition task with 2^21.5 queries in under an hour, such that the extracted model agrees with the oracle on all inputs up to a worst-case error of 2^-25, or a model with 4,000 parameters in 2^18.5 queries with worst-case error of 2^-40.4. Code is available at https://github.com/google-research/cryptanalytic-model-extraction.