This paper presents a device-independent security proof for Quantum Key Distribution (QKD) against collective attacks. The security analysis is based on the violation of Bell-type inequalities and does not require any assumptions about the internal workings of the QKD devices. The main result is a tight bound on the Holevo information between one of the authorized parties and the eavesdropper, as a function of the amount of violation of a Bell-type inequality. QKD allows two parties, Alice and Bob, to generate a secret key in the presence of an eavesdropper, Eve. The secret key can then be used for encryption. The first QKD protocol, BB84, was introduced by Bennett and Brassard in 1984. Since then, QKD has seen significant theoretical and experimental advances. Existing QKD schemes rely on several assumptions, including that any eavesdropper must obey the laws of quantum physics, and that the measurement settings and outcomes are secret. These assumptions ensure that no unwanted classical information leaks from Alice's and Bob's laboratories. However, existing security proofs assume that Alice and Bob have perfect control over the state preparation and measurement devices, which is often critical for the security of the protocol. This paper presents a device-independent security proof against collective attacks by a quantum Eve for the protocol described in Ref. [9]. The proof holds under no other requirements than the essential ones listed above. It is therefore "device-independent" in the sense that it needs no knowledge of the way the QKD devices work, provided quantum physics is correct and provided Alice and Bob do not allow any unwanted signal to escape from their laboratories. In a collective attack, Eve applies the same attack on each particle of Alice and Bob. The physical basis for the device-independent security proof is the fact that measurements on entangled particles can provide Alice and Bob with non-local correlations, i.e., correlations that cannot be reproduced by shared randomness (local variables), as detected by the violation of Bell-type inequalities. The intuition that Eve cannot have full information about these correlations is supported by the fact that the violation of Bell-type inequalities implies that the correlations are non-local. The paper presents a protocol that is a modification of the Ekert 1992 protocol. Alice and Bob share a quantum channel consisting of a source that emits pairs of entangled particles. On each of her particles, Alice chooses between three possible measurements, and Bob between two possible measurements. The raw key is extracted from the pair {A0, B1}. The quantum bit error rate (QBER) is Q = prob(a0 ≠ b1). The protocol uses the measurements A1, A2, B1, and B2 on a subset of their particles to compute the CHSH polynomial, which defines the CHSH inequality S ≤ 2. The paper also presents an upper-bound on the Holevo quantity, which is the main result. The bound is derived using a theorem that relates the HolevoThis paper presents a device-independent security proof for Quantum Key Distribution (QKD) against collective attacks. The security analysis is based on the violation of Bell-type inequalities and does not require any assumptions about the internal workings of the QKD devices. The main result is a tight bound on the Holevo information between one of the authorized parties and the eavesdropper, as a function of the amount of violation of a Bell-type inequality. QKD allows two parties, Alice and Bob, to generate a secret key in the presence of an eavesdropper, Eve. The secret key can then be used for encryption. The first QKD protocol, BB84, was introduced by Bennett and Brassard in 1984. Since then, QKD has seen significant theoretical and experimental advances. Existing QKD schemes rely on several assumptions, including that any eavesdropper must obey the laws of quantum physics, and that the measurement settings and outcomes are secret. These assumptions ensure that no unwanted classical information leaks from Alice's and Bob's laboratories. However, existing security proofs assume that Alice and Bob have perfect control over the state preparation and measurement devices, which is often critical for the security of the protocol. This paper presents a device-independent security proof against collective attacks by a quantum Eve for the protocol described in Ref. [9]. The proof holds under no other requirements than the essential ones listed above. It is therefore "device-independent" in the sense that it needs no knowledge of the way the QKD devices work, provided quantum physics is correct and provided Alice and Bob do not allow any unwanted signal to escape from their laboratories. In a collective attack, Eve applies the same attack on each particle of Alice and Bob. The physical basis for the device-independent security proof is the fact that measurements on entangled particles can provide Alice and Bob with non-local correlations, i.e., correlations that cannot be reproduced by shared randomness (local variables), as detected by the violation of Bell-type inequalities. The intuition that Eve cannot have full information about these correlations is supported by the fact that the violation of Bell-type inequalities implies that the correlations are non-local. The paper presents a protocol that is a modification of the Ekert 1992 protocol. Alice and Bob share a quantum channel consisting of a source that emits pairs of entangled particles. On each of her particles, Alice chooses between three possible measurements, and Bob between two possible measurements. The raw key is extracted from the pair {A0, B1}. The quantum bit error rate (QBER) is Q = prob(a0 ≠ b1). The protocol uses the measurements A1, A2, B1, and B2 on a subset of their particles to compute the CHSH polynomial, which defines the CHSH inequality S ≤ 2. The paper also presents an upper-bound on the Holevo quantity, which is the main result. The bound is derived using a theorem that relates the Holevo